[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16185169#comment-16185169
]
Hyukjin Kwon commented on SPARK-21190:
--
Will keep in mind and suggest to fix it or fix it by myself when we happen to
fix some codes around here (of course before the release).
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16185166#comment-16185166
]
Reynold Xin commented on SPARK-21190:
-
OK it would be great to have a better error message, e.g. remove "currently",
and tell users to work around the limitation, they can create an 1-arg udf and
ignore the arg.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16185164#comment-16185164
]
Hyukjin Kwon commented on SPARK-21190:
--
[~rxin], I suggested to disallow it
[here|https://github.com/apache/spark/pull/18659#discussion_r139856165] and I
think reviewers and committers agreed with it for the similar reasons.
So, it was separately fixed in [this
commit|https://github.com/apache/spark/commit/d8e825e3bc5fdb8ba00eba431512fa7f771417f1].
Therefore, the cases below are not allowed:
{code}
>>> from pyspark.sql.functions import pandas_udf
>>> from pyspark.sql.types import *
>>> @pandas_udf(returnType=LongType())
... def add_one():
... return 1
...
Traceback (most recent call last):
File "", line 1, in
File "/.../spark/python/pyspark/sql/functions.py", line 2133, in _udf
raise NotImplementedError("0-parameter pandas_udfs are not currently
supported")
NotImplementedError: 0-parameter pandas_udfs are not currently supported
>>> pandas_udf(lambda: 1, "long")
Traceback (most recent call last):
File "", line 1, in
File "/.../spark/python/pyspark/sql/functions.py", line 2209, in pandas_udf
wrapped_udf = _create_udf(f, returnType=returnType, vectorized=True)
File "/.../spark/python/pyspark/sql/functions.py", line 2144, in _create_udf
return _udf(f=f, returnType=returnType, vectorized=vectorized)
File "/.../spark/python/pyspark/sql/functions.py", line 2133, in _udf
raise NotImplementedError("0-parameter pandas_udfs are not currently
supported")
NotImplementedError: 0-parameter pandas_udfs are not currently supported
>>> pandas_udf(lambda: 1, LongType())
Traceback (most recent call last):
File "", line 1, in
File "/.../spark/python/pyspark/sql/functions.py", line 2209, in pandas_udf
wrapped_udf = _create_udf(f, returnType=returnType, vectorized=True)
File "/.../spark/python/pyspark/sql/functions.py", line 2144, in _create_udf
return _udf(f=f, returnType=returnType, vectorized=vectorized)
File "/.../spark/python/pyspark/sql/functions.py", line 2133, in _udf
raise NotImplementedError("0-parameter pandas_udfs are not currently
supported")
NotImplementedError: 0-parameter pandas_udfs are not currently supported
{code}
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing thi
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16185155#comment-16185155
]
Reynold Xin commented on SPARK-21190:
-
Where did we settle on 0-arg UDFs? I think we should just disallow it, since
users can trivially workaround it by creating a 1-arg UDF that ignores the arg.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16177886#comment-16177886
]
Reynold Xin commented on SPARK-21190:
-
Maybe create an umbrella ticket so it is easier to link.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16177885#comment-16177885
]
Wenchen Fan commented on SPARK-21190:
-
yea, let's do that in a separated ticket.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16177846#comment-16177846
]
Li Jin commented on SPARK-21190:
[~cloud_fan], do we want to track other vectorized udf efforts (group, window,
etc) in this Jira or a separate one?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Bryan Cutler
> Labels: SPIP
> Fix For: 2.3.0
>
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16162646#comment-16162646
]
Takuya Ueshin commented on SPARK-21190:
---
[~icexelloss] Thank you for your suggestion. I agree that we should unify the
prs to 1 pr and the collaboration.
As [~cloud_fan] mentioned in [~bryanc]'s pr, I'll send some prs toward
[~bryanc]'s branch per functionality to discuss each.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: i
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16162205#comment-16162205
]
Bryan Cutler commented on SPARK-21190:
--
Thanks [~icexelloss]. I definitely think collaboration should be encouraged,
especially in this case where there was existing work done by you and I that
could have been leveraged. The major difference between
[#18659|https://github.com/apache/spark/pull/18659] and
[#19147|https://github.com/apache/spark/pull/19147] is the Arrow data format
used. #19147 uses the Arrow stream format which has some pros and cons that I
brought up in the PR.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This messag
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16158845#comment-16158845
]
Li Jin commented on SPARK-21190:
To [~bryanc]'s point, PR [#18659|https://github.com/apache/spark/pull/18659]
and PR [#19147|https://github.com/apache/spark/pull/19147] are largely similar
and it makes sense not to keep two PR for the same thing.
Also, what I am curious about is what kind of guide lines to follow to avoid
such duplicate work in the future. IMHO, [~bryanc] has been linked his PR to
this Jira a while back and has been actively engaging in all discussions, so I
am not sure why do we need a similar second PR in this case. (Of course,
however, if people think #18659 and #19147 are very different, that's another
story).
I have worked with [~bryanc] in the past together on SPARK-13534 and
collaborating on the same branch worked quite well for us. Maybe that's
something we should encourage?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16156082#comment-16156082
]
Bryan Cutler commented on SPARK-21190:
--
I attached my PR because it had already been done and pretty much matches the
proposal from here. It is slightly different to define the UDF, I use just a
flag in the normal UDF declaration, for example {{udf(my_func, DoubleType(),
vectorized=True}}, but it is trivial to change it to {{pandas_udf}} if that is
what's decided. I'm not exactly sure if there was a problem in using my PR for
this, but I'm more than happy to discuss the differences and hopefully avoid
duplicated work.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16156068#comment-16156068
]
Apache Spark commented on SPARK-21190:
--
User 'BryanCutler' has created a pull request for this issue:
https://github.com/apache/spark/pull/18659
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16155271#comment-16155271
]
Apache Spark commented on SPARK-21190:
--
User 'ueshin' has created a pull request for this issue:
https://github.com/apache/spark/pull/19147
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16155270#comment-16155270
]
Takuya Ueshin commented on SPARK-21190:
---
[~leif], [~bryanc] Thanks for the instruction about {{inspect}} module.
I believe we got a consensus about the basic APIs except for the size hint, so
I've decided to submit a pr based on my proposal at first and let's discuss the
size hint for 0-parameter UDF (or for more parameter UDFs to be consistent)
through the review process.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
--
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16154361#comment-16154361
]
Bryan Cutler commented on SPARK-21190:
--
Thanks [~ueshin], I think having an optional {{kwargs}} at least would make
things consistent across all types of {{pandas_udf}}. Here is how you could
use to check that a function has a {{**}} arg that would accept keyword
arguments:
{noformat}
In [1]: import inspect
In [2]: def f(a, b, **kwargs):
...: print(a, b, kwargs)
In [3]: def g(a, b):
...: print(a, b)
In [4]: inspect.getargspec(f)
Out[4]: ArgSpec(args=['a', 'b'], varargs=None, keywords='kwargs', defaults=None)
In [5]: inspect.getargspec(g)
Out[5]: ArgSpec(args=['a', 'b'], varargs=None, keywords=None, defaults=None)
In [6]: if inspect.getargspec(f).keywords is not None:
...: f(1, 2, size=3)
...:
(1, 2, {'size': 3})
{noformat}
If the user defines a 0-parameter UDF without a {{**kwargs}} then it is
probably best to raise an error if the returned size isn't right - although it
would be possible to repeat and/or slice..
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16153986#comment-16153986
]
Leif Walsh commented on SPARK-21190:
I think the size parameter is confusing: if a 1-or-more-parameter UDF gets
called with some number of Series or DataFrame objects, but a 0-parameter UDF
switches to getting a size scalar, it's less consistent. In fact, the most
consistent interface would be for all UDFs to get called with a single
DataFrame parameter, containing all the columns, already aligned. However, I
somewhat more like the idea of a UDF on multiple columns getting separate
Series objects, as this makes the UDF's function signature look more similar to
how it will be called.
You can use the {{inspect}} module in the python standard library to find out
what parameters a function accepts, but I'd like to warn you against using a
magic {{**kwargs}} parameter, that's a fairly non-intuitive API and is also
somewhat brittle.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16153978#comment-16153978
]
Takuya Ueshin commented on SPARK-21190:
---
[~leif] Thank you for your proposal.
I'm sorry but I couldn't figure out the big difference between only size
parameter and a Series/DataFrame with no content and an index, and I guess
no-content Series/DataFrame confuses users more.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issue
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16153955#comment-16153955
]
Takuya Ueshin commented on SPARK-21190:
---
[~bryanc] Thank you for your suggestion. {{**kwargs}} might be a good idea to
provide the {{size}} hint and other metadata in the future. I'm not sure how to
inspect the UDF to check if it accepts kwargs yet, though.
Should we make {{**kwargs}} required? If not, users can still define
0-parameter UDF without it and we can't determine how to handle it.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16151377#comment-16151377
]
Leif Walsh commented on SPARK-21190:
You can also make a Series with no content and an index, but this becomes a
float64 Series full of NaNs, which seems less good: it's indistinguishable from
a column which is actually full of NaNs:
{noformat}
In [5]: pd.Series(index=list(range(100)))
Out[5]:
0NaN
1NaN
2NaN
3NaN
4NaN
5NaN
6NaN
7NaN
8NaN
9NaN
10 NaN
11 NaN
12 NaN
13 NaN
14 NaN
15 NaN
16 NaN
17 NaN
18 NaN
19 NaN
20 NaN
21 NaN
22 NaN
23 NaN
24 NaN
25 NaN
26 NaN
27 NaN
28 NaN
29 NaN
..
70 NaN
71 NaN
72 NaN
73 NaN
74 NaN
75 NaN
76 NaN
77 NaN
78 NaN
79 NaN
80 NaN
81 NaN
82 NaN
83 NaN
84 NaN
85 NaN
86 NaN
87 NaN
88 NaN
89 NaN
90 NaN
91 NaN
92 NaN
93 NaN
94 NaN
95 NaN
96 NaN
97 NaN
98 NaN
99 NaN
Length: 100, dtype: float64
{noformat}
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16151376#comment-16151376
]
Leif Walsh commented on SPARK-21190:
Yep, that's totally a thing:
{noformat}In [1]: import pandas as pd
In [2]: pd.DataFrame(index=list(range(100)))
Out[2]:
Empty DataFrame
Columns: []
Index: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39,
40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59,
60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]
[100 rows x 0 columns]
In [3]: df = Out[2]
In [4]: pd.Series(data=1, index=df.index)
Out[4]:
0 1
1 1
2 1
3 1
4 1
5 1
6 1
7 1
8 1
9 1
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
..
701
711
721
731
741
751
761
771
781
791
801
811
821
831
841
851
861
871
881
891
901
911
921
931
941
951
961
971
981
991
Length: 100, dtype: int64
{noformat}
So, how about:
{noformat}
@pandas_udf(LongType())
def f0(df):
return pd.Series(data=1, index=df.index)
df.select(f0())
{noformat}
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one co
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16151261#comment-16151261
]
Leif Walsh commented on SPARK-21190:
I'm not 100% sure this is legal pandas but I think it might be. If no columns
are passed in, we might be able to pass in a zero-column dataframe with an
integer index from 0 to size-1. Then {{of.Series(data=1, index=D.C.index)}}
might do the right thing, without magic extra parameters. I think this would be
conceptually correct too. I will verify and provide sample code this weekend if
I find the time.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16150924#comment-16150924
]
Bryan Cutler commented on SPARK-21190:
--
I'm good with the API summary proposed by [~ueshin], but I'm also not crazy
about the 0-parameter UDF. My only other idea would be to allow any
`pandas_udf` to optionally define `**kwargs` as the last argument in the UDF.
In the python worker, it would be easy to inspect the UDF to check if it
accepts kwargs and then provide the `size` hint (could add other metadata
also). Of course isn't the most intuitive also, but at least would be
consistent across all `pandas_udf`s. So a 0-parameter example would look like:
{code}
@pandas_udf(LongType())
def f0(**kwargs):
return pd.Series(1).repeat(kwargs['size'])
df.select(f0())
{code}
It's not perfect since it would require the user to know that size is in the
kwargs, but just thought I'd throw it out as an alternative.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Desig
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16143491#comment-16143491
]
Wenchen Fan commented on SPARK-21190:
-
hmmm, your proposal has a weird usage: users need to pass an argument to a
0-parameter UDF, and transferring this constant column batch is also a waste.
Defining the 0-parameter udf with a size parameter is unintuitive, but I can't
think of a better idea...
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional co
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16140094#comment-16140094
]
Li Jin commented on SPARK-21190:
[~ueshin], Got it. I'd actually prefer doing it this way:
{code}
@pandas_udf(LongType())
def f0(v):
return pd.Series(1).repeat(len(v))
df.select(f0(F.lit(0)))
{code}
Instead of passing the size as a scalar to the function.
Passing size to f0 feels intuitive to me because f0 is defined as a function
that takes one argument - size, but being invoked with 0 args - f0().
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16139507#comment-16139507
]
Takuya Ueshin commented on SPARK-21190:
---
[~icexelloss]
We can know the length of input from {{RecordBatch}} and pass it as a hint to
the actual function for users to create the output of the exact length.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16138340#comment-16138340
]
Li Jin commented on SPARK-21190:
[~ueshin], thanks for the summary. +1 for this API.
Although the 0-parameter UDF is a bit confusing. Where does "size" come from?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16137897#comment-16137897
]
Takuya Ueshin commented on SPARK-21190:
---
Hi all,
I'd like to summarize this discussion about APIs for vectorized UDFs (not for
UDAFs or Window operations for now).
Please correct me if I miss anything or add comments if you have some.
*Proposed API*
Definition of Vectorized UDFs
We introduce a {{@pandas_udf}} decorator (or annotation) to define vectorized
UDFs which takes one or more pandas.Series or one integer value meaning the
length of the input value for 0-parameter UDFs. The return value should be
pandas.Series of the specified type and the length of the returned value should
be the same as input value.
We can define vectorized UDFs as:
{noformat}
@pandas_udf(DoubleType())
def plus(v1, v2):
return v1 + v2
{noformat}
or we can define as:
{noformat}
plus = pandas_udf(lambda v1, v2: v1 + v2, DoubleType())
{noformat}
We can use it similar to row-by-row UDFs:
{noformat}
df.withColumn('sum', plus(df.v1, df.v2))
{noformat}
As for 0-parameter UDFs, we can define and use as:
{noformat}
@pandas_udf(LongType())
def f0(size):
return pd.Series(1).repeat(size)
df.select(f0())
{noformat}
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16105095#comment-16105095
]
Li Jin commented on SPARK-21190:
I think the use case 2 of what [~rxin] proposed originally is a good API to
enable first. I think it can a bit better if the input of the user function is
not a {{pandas.DataFrame}} but {{pandas.Series}} to match Spark columns. i.e.,
instead of:
{code}
@spark_udf(some way to describe the return schema)
def my_func(input):
""" Some user-defined function.
:param input: A Pandas DataFrame with two columns, a and b.
:return: :class: A numpy array
"""
return input[a] + input[b]
df = spark.range(1000).selectExpr("id a", "id / 2 b")
df.withColumn("c", my_func(df.a, df.b))
{code}
I think this is better:
{code}
@spark_udf(some way to describe the return schema)
def my_func(a, b):
""" Some user-defined function.
:param input: Two Pandas Series, a and b
:return: :class: A Pandas Series
"""
return a + b
df = spark.range(1000).selectExpr("id a", "id / 2 b")
df.withColumn("c", my_func(df.a, df.b))
{code}
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
>
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16105072#comment-16105072
]
Li Jin commented on SPARK-21190:
[~cloud_fan], thanks for pointing out `ArrowColumnVector`. [~bryanc], I think
#18659 could serve as a basis for future udf work. My work with #18732 has some
overlap with #18659 but I can work with [~bryanc] to merge.
[~cloud_fan] and [~rxin], do you have chance to think more about the API?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apa
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16102420#comment-16102420
]
Bryan Cutler commented on SPARK-21190:
--
Hi [~icexelloss], yes I think there is definitely some basic framework needed
to enable vectorized UDFs in Python that could be done before any API
improvements. Like you pointed out in (3) the PythonRDD and pyspark.worker
will need to be more flexible to handle a different udf, as well as some work
on the pyspark.serializer to produce/consume vector data. I was hoping
[#18659|https://github.com/apache/spark/pull/18659] could serve as this basis.
I plan on updating it to use the {{ArrowColumnVector}} and {{ArrowWriter}} from
[~ueshin] when that gets merged.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16101014#comment-16101014
]
Wenchen Fan commented on SPARK-21190:
-
I think (2) is already done by {{ArrowColumnVector}} (written by [~ueshin]),
and you can just call `ColumnarBatch.rowIterator` for the conversion.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16100712#comment-16100712
]
Li Jin commented on SPARK-21190:
[~bryanc],
I have looked at your PR at https://github.com/apache/spark/pull/18659 and we
have quite bit of overlap in (2) and (3) above. These are not binded to the API
choices and I think we should combine/unify them and establish the base line.
What do you think?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional com
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16100582#comment-16100582
]
Li Jin commented on SPARK-21190:
I have created this PR for the groupby().apply() use case with pandas udf:
https://github.com/apache/spark/pull/18732
This PR consists of a few parts that can be broken into smaller PRs:
(1) pandas_udf:
This is a new API in pyspark.sql.functions what allows using pandas udf with
various pyspark sql functions discussed here
(2) arrow batch record -> unsafe row conversion
This is a set of converters to turn arrow batch record -> unsafe rows. (The
other direction has already been implemented). This is needed for pandas udf
independent of the API choices.
(3) df.groupby().apply() function in pyspark sql
This is a new API in pyspark sql that implements split-apply-merge pattern with
pyspark and pandas udf.
I am hoping to get some feedbacks here as well as working together to figure
the next step for vectorized UDFs
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16083018#comment-16083018
]
Bryan Cutler commented on SPARK-21190:
--
[~cloud_fan] yes, I know not every function can be vectorized so I was saying
it would need an SQLConf to enable, or maybe something less broad. The format
in my example is pretty standard for writing vectorized fuctions (at least from
my experience with numpy/matlab), and I think a lot of users probably have UDFs
that fit into this mold. It would be a great benefit to Python users to be
able to use vectorization without having to change their existing UDFs.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16079954#comment-16079954
]
Wenchen Fan commented on SPARK-21190:
-
[~bryanc] You example works because Python is dynamically typed and {{a + b}}
can work for both scalar values and vector values. However I'm afraid this may
not apply for all operators/functions, and I think it's better to introduce new
APIs so that users can know they are running scalar udf or columnar udf.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: i
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16078837#comment-16078837
]
Bryan Cutler commented on SPARK-21190:
--
[~rxin] I was talking about 2 different things, sorry if I was unclear. My
first point was just to say that we could get vectorized UDFs for the case of
returning 1 column with the existing pyspark udf api. Just a small change to
the function in your use case 1:
{noformat}
def my_func_that_returns_one_column(a, b):
""" Some user-defined function.
:param input: a and b, scalar or numpy array.
:return: scalar or numpy array
"""
return a + b
{noformat}
This function is vectorized and can be composed as a standard UDF, specifying
the return type
{noformat}
my_func = udf(my_func_that_returns_one_column, DoubleType())
df = spark.range(1000).selectExpr("id a", "id / 2 b")
df.withColumn("c", my_func(df.a, df.b))
{noformat}
The python worker would then load the Arrow record batch and extract columns
"a" and "b" as pandas.Series and apply the function, basically like this
{noformat}
sa, sb # columns "a" and "b" read by Arrow as pandas.Series
sc = f(sa, sb) # where f is my_func_that_returns_one_column and sc is a
pandas.Series with same size as "sa" and "sb"
{noformat}
So if this gives us vectorized UDFs for 1 column, I'm not sure what exactly
expanding the API will buy us?
btw, I'm not against any of what has been proposed here, I'm just bringing this
up to make sure it's known and facilitate discussion.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16077484#comment-16077484
]
Reynold Xin commented on SPARK-21190:
-
[~bryanc] Sorry I don't think it makes sense to not introduce APIs. In general
we are trying to raise the level of abstractions and move people away from
lower level APIs. We wouldn't want to give users another reason to use the
lower level RDD API.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> A few things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
> 3. How do we handle null values, since Pandas doesn't have the concept of
> nulls?
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands,
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16077390#comment-16077390
]
Bryan Cutler commented on SPARK-21190:
--
This is a great discussion so far and I would love to see this in Spark as a
future use of Arrow! Thanks for the detailed use cases [~icexelloss] and
[~leif]. I've also been thinking about this optimization for a while and have
been working out some POCs to experiment with. Here are my some of my thoughts
on the discussion so far
1. Apply a vectorized UDF that returns 1 column of same size
I think this can be accomplished within the current API for declaring UDFs,
only using Arrow on the back-end that will load the columns as Pandas.Series
and return a Pandas.Series. It might be necessary to do some conversion on the
returned Pandas.Series to ensure it is the defined return type and NULL values
are correct, but all this is known so I don't think it's necessary to add extra
annotation to the UDF. This functionality could be enabled with a SQLConf.
2. Apply a vectorized UDAF - windowing use case in mind that aggregates to a
scaler
This would be great to have, but seems like it could be quite a bit more
complicated than a standard UDF so I agree with [~rxin] that this is probably
better to enable in the future.
3. Apply a function to a Pandas DataFrame that returns a new Pandas DataFrame
I think it is a little awkward to use Pandas DataFrames in the same way as UDFs
if it's required to define all columns types in the returned frame, and also
not very flexible. Here is an alternative that I think could work for some of
the use cases described above:
*Proposed API*
This would require some light wrappers around a few pyspark.sql classes. A
method {{DataFrame.asPandas()}} could be added which would produce a wrapped
DataFrame and allow processing with Pandas.DataFrames. See here for a rough
design https://gist.github.com/BryanCutler/2d2ae04e81fa96ba4b61dc095726419f
Use Case 1. Apply a function that takes a pandas.DataFrame as input and
produces a new pandas.DataFrame. Similar to use case 1 above, maybe not that
interesting for data analysis, but usage is straight-forward
{noformat}
def my_func_on_entire_df(input):
""" Some user-defined function.
:param input: A Pandas DataFrame with two columns, a and b.
:return: :class: A Pandas data frame.
"""
input[c] = input[a] input[b]
input[d] = input[a] - input[b]
return input
spark.range(1000).selectExpr("id a", "id / 2 b")
.asPandas().rdd.map(my_func_on_entire_df)
{noformat}
This will change the pyspark.DataFrame to an wrapped RDD that uses Arrow to
serde and defines methods to map/reduce/collect with functions accepting
pandas.DataFrames as input/output.
Use Case 2. Calculate the weighted mean on a column in a Pandas.DataFrame
{noformat}
def weighted_mean(df):
"""Take weighted mean of a column.
:param df: A Pandas DataFrame with data column "v" and weight column "w"
:return: :class: A Pandas data frame with column 'wmu'.
"""
df['wmu'] = np.average(input['v'], weights=input['w'])
return df
spark.range(1000).selectExpr("id % 7 a", "id / 2 v", "id % 2 w")
.asPandas().groupBy("a").agg(weighted_mean)
{noformat}
Here a subclass of GroupedData is created that overrides the "agg()" method to
accept a function using pandas.DataFrame. The "agg()" method would convert the
resulting pandas.DataFrames to Arrow and then convert Arrow to a Spark Dataset.
Using this to apply over a window is a little more complicated, so I'll have to
think more on that.
The benefit of this design is that it only relies on conversions to/from Arrow,
which has a well defined schema, so it's easy to compose functions and work
with them. However, the downside is that it's mostly separated from the rest
of SparkSQL and doesn't fit in sql expressions like UDFs. So this might lead
to having to do extra work, like join DataFrames, after the Pandas processing
is complete.
Regardless, I think vectorized UDFs without any API change would be a great
place to start and allow us to optimize the processing with Arrow before
getting into more complex usage.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serializ
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16074857#comment-16074857
]
Leif Walsh commented on SPARK-21190:
If the user specifies an int return type but produces floats in order to
indicate nulls, the "right thing" would be to convert to nullable int. If the
user specifies int return type but produces float because of operator error,
the "right thing" would be to produce an exception with a meaningful error
message. Unfortunately I do not have a way to distinguish.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
F
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16074848#comment-16074848
]
Li Jin commented on SPARK-21190:
> I have 2 thoughts:
> 1. How should we handle null values? IIRC, pd.Series doesn't work well with
> null, e.g. int series with null becomes float type, boolean series with null
> becomes object type.
> 2. I think UDF returning scalar value only makes sense for group/window,
> otheriwse the result is indeterministic for users.
Here are my thoughts:
1. From Spark -> Pandas, I think turning nullable int series to float series
is fine because pandas users are already dealing with issue in pure pandas.
From Pandas -> Spark, because Arrow supports nullable type, we can turn
the pandas series into arrow series with the correct type. For instance, if the
user specifies that udf returns a int series but ends
up returning a float series because it contains nullable values, we can
turn the pandas float series into arrow nullable int series. This might have
some performance cost but hopefully is not too bad. [~wesmckinn], do you have
some thoughts on this?
2. You are absolutely right, not every return types makes sense with every
function. I think it's fine we can do some type checking in catalyst and throw
exceptions.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :r
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16072429#comment-16072429
]
Leif Walsh commented on SPARK-21190:
I believe we could also compute window indexes while we stream over the data
and serialize to arrow, don't need to keep the data in the jvm necessarily.
I'm more worried about keeping a whole partition in python. You could have
small windows but large partitions and pay an unfair memory cost unless we have
some smart heuristics about how to batch the data into sub-partition chunks.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubs
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16072424#comment-16072424
]
Leif Walsh commented on SPARK-21190:
I figure we could address that by using shared memory, if we need to. Don't
think that's a blocker for a first cut though, just my opinion.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16072019#comment-16072019
]
Wenchen Fan commented on SPARK-21190:
-
> I think we can get away with doing windowing (deciding which rows to
> aggregate) in scala to avoid reimplementing that logic in python, and then
> send row indexes (a list of begin/end pairs which each are the bounds of a
> window) over to python to select the windows.
Does this mean we need to keep the data of the whole window in both JVM and
Python? This seems too expensive...
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsub
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16071002#comment-16071002
]
Wenchen Fan commented on SPARK-21190:
-
Thanks for your proposal!
I have 2 thoughts:
1. How should we handle null values? IIRC, {{pd.Series}} doesn't work well with
null, e.g. int series with null becomes float type, boolean series with null
becomes object type.
2. I think UDF returning scalar value only makes sense for group/window,
otheriwse the result is indeterministic for users.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16070809#comment-16070809
]
Leif Walsh commented on SPARK-21190:
I think we can get away with doing windowing (deciding which rows to aggregate)
in scala to avoid reimplementing that logic in python, and then send row
indexes (a list of begin/end pairs which each are the bounds of a window) over
to python to select the windows.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16070618#comment-16070618
]
Li Jin commented on SPARK-21190:
I have some APIs design written down here:
Here is how to define a udf:
https://github.com/icexelloss/spark/blob/pandas-udf-doc/docs/pyspark-pandas-udf.md#definition-of-pandas-udf
Here is example how to use udf:
1. Vectorized row transforms:
https://github.com/icexelloss/spark/blob/pandas-udf-doc/docs/pyspark-pandas-udf.md#vectorized-row-operations
2. Window:
https://github.com/icexelloss/spark/blob/pandas-udf-doc/docs/pyspark-pandas-udf.md#window-operations
3. Group:
https://github.com/icexelloss/spark/blob/pandas-udf-doc/docs/pyspark-pandas-udf.md#group-operations
To give a quick an example (for group aggregations):
{code:python}
import numpy as np
@pandas_udf(DoubleType())
def weighted_mean_udf(v, w):
return np.average(v, weights=w)
df.groupBy('id').agg(weighted_mean_udf(df.v1, df.w).as('v1_wm'))
{code}
In this example, the udf takes one or more pandas.Series of the same size, and
returns a scalar value. The result is the aggregation for each group.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16069665#comment-16069665
]
Wenchen Fan commented on SPARK-21190:
-
For aggregate, I think it makes more sense to do the grouping at Spark side,
and then send each group to Python, and use Pandas to do reduce or add extra
column. We can have a config for group max size, and when sending groups from
JVM to Python, fail the query if the max size is exceeded.
For window, like [~leif] said, ideally we should not send each window to
python, to avoid transferring a lot of duplicate data. Instead, we can send the
entire partition to python, in streaming manner. Then we do the windowing at
python side and use Pandas to do reduce or something. The downside is we need
to re-implement many things in Python.
I'll think about the API too.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same im
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16069490#comment-16069490
]
Reynold Xin commented on SPARK-21190:
-
That makes a lot of sense. So to design APIs similar to a lot of the ply
operations. I will think a bit about this next week. Btw if somebody can also
just propose some different alternative APIs (perhaps as a gist), it'd
facilitate the discussion further.
How should we do memory management though? Should we just have a configurable
cap on the maximum number of roles? Should we just let user crash (not a great
experience)?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
--
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16069269#comment-16069269
]
Leif Walsh commented on SPARK-21190:
I agree with [~icexelloss] that we should aim to provide an API which provides
"logical" groups of data to the UDF rather than the implementation detail of
providing partitions wholesale. Setting aside for the moment the problem with
dataset skew which could cause one group to be very large, let's look at some
use cases.
One obvious use case that tools like dplyr and pandas support is
{{df.groupby(...).aggregate(...)}}. Here, we group on some key and apply a
function to each logical group. This can be used to e.g. demean each group
w.r.t. its cohort. Another use case that we care about with
[Flint|https://github.com/twosigma/flint] is aggregating over a window. In
pandas terminology this is the {{rolling}} operator. One might want to, for
each row, perform a moving window average or rolling regression over a history
of some size. The windowed aggregation poses a performance question that the
groupby case doesn't: namely, if we naively send each window to the python
worker independently, we're transferring a lot of duplicate data since each
overlapped window contains many of the same rows. An option here is to
transfer the entire partition on the backend and then instruct the python
worker to call the UDF with slices of the whole dataset according to the
windowing requested by the user.
I think the idea of presenting a whole partition in a pandas dataframe to a UDF
is a bit off-track. If someone really wants to apply a python function to the
"whole" dataset, they'd be best served by pulling those data back to the driver
and just using pandas, if they tried to use spark's partitions they'd get
somewhat arbitrary partitions and have to implement some kind of merge operator
on their own. However, with grouped and windowed aggregations, we can provide
an API which truly is parallelizable and useful.
I want to focus on use cases where we actually can parallelize without
requiring a merge operator right now. Aggregators in pandas and related tools
in the ecosystem usually assume they have access to all the data for an
operation and don't need to merge results of subaggregations. For aggregations
over larger datasets you'd really want to encourage the use of native Spark
operations (that use e.g. {{treeAggregate}}).
Does that make sense? I think it focuses the problem nicely that it becomes
fairly tractable.
I think the really hard part of this API design is deciding what the inputs and
outputs of the UDF look like, and providing for the myriad use cases therein.
For example, one might want to aggregate each group down to a scalar (e.g.
mean) and do something with that (either produce a reduced dataset with one
value per group, or add a column where each group has the same value across all
rows), or one might want to compute over the group and produce a value per row
within the group and attach that as a new column (e.g. demeaning or ranking).
These translate roughly to the differences between the [**ply operations in
dplyr|https://www.jstatsoft.org/article/view/v040i01/v40i01.pdf] or the
differences in pandas between {{df.groupby(...).agg(...)}} and
{{df.groupby(...).transform(...)}} and {{df.groupby(...).apply(...)}}.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16063873#comment-16063873
]
Li Jin commented on SPARK-21190:
[~r...@databricks.com],
The use case of seeing entire partition at a time is one of the less
interesting use case, because partition is somehow an implementation detail of
spark. From the use cases I saw from our internal pyspark/pandas users, the
majority interests are to use pandas udf with a spark windowing and grouping
operation, for instance, breaking data into day/month/year piece and perform
some analysis using pandas for each piece.
However, the challenge still holds if one group is too large. One idea is to
implement local aggregation and merge in pandas operations so it can be
executed efficiently.
I think without solving the problem of "one group doesn't fit in memory", this
can already be a pretty powerful tool because currently to perform any pandas
analysis, the *entire dataset* has to fit in memory (that is, without using
tools like dask).
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16063515#comment-16063515
]
Reynold Xin commented on SPARK-21190:
-
[~icexelloss] Thanks. Your proposal brings up a good point, which is do we want
to allow users to see the entire partition at a time. As demonstrated by your
example, this can be a very powerful tool (especially for leveraging the time
series analysis features available in Pandas). The primary challenge there is
that a partition can be very large, and it is possible to run out of memory
when using Pandas. Any thoughts on how we should handle those cases?
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
[jira] [Commented] (SPARK-21190) SPIP: Vectorized UDFs in Python
[
https://issues.apache.org/jira/browse/SPARK-21190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16063103#comment-16063103
]
Li Jin commented on SPARK-21190:
Very excited to see this.
I created https://issues.apache.org/jira/browse/SPARK-20396 earlier and I have
been working the same problem for a while. [~julienledem] and I talked about
the work we did on Spark Summit this year.
Here is the design doc:
https://github.com/icexelloss/spark/blob/pandas-udf-doc/docs/pyspark-pandas-udf.md
I think the ideas are very similar and we could probably collaborate.
> SPIP: Vectorized UDFs in Python
> ---
>
> Key: SPARK-21190
> URL: https://issues.apache.org/jira/browse/SPARK-21190
> Project: Spark
> Issue Type: New Feature
> Components: PySpark, SQL
>Affects Versions: 2.2.0
>Reporter: Reynold Xin
>Assignee: Reynold Xin
> Labels: SPIP
> Attachments: SPIPVectorizedUDFsforPython (1).pdf
>
>
> *Background and Motivation*
> Python is one of the most popular programming languages among Spark users.
> Spark currently exposes a row-at-a-time interface for defining and executing
> user-defined functions (UDFs). This introduces high overhead in serialization
> and deserialization, and also makes it difficult to leverage Python libraries
> (e.g. numpy, Pandas) that are written in native code.
>
> This proposal advocates introducing new APIs to support vectorized UDFs in
> Python, in which a block of data is transferred over to Python in some
> columnar format for execution.
>
>
> *Target Personas*
> Data scientists, data engineers, library developers.
>
> *Goals*
> - Support vectorized UDFs that apply on chunks of the data frame
> - Low system overhead: Substantially reduce serialization and deserialization
> overhead when compared with row-at-a-time interface
> - UDF performance: Enable users to leverage native libraries in Python (e.g.
> numpy, Pandas) for data manipulation in these UDFs
>
> *Non-Goals*
> The following are explicitly out of scope for the current SPIP, and should be
> done in future SPIPs. Nonetheless, it would be good to consider these future
> use cases during API design, so we can achieve some consistency when rolling
> out new APIs.
>
> - Define block oriented UDFs in other languages (that are not Python).
> - Define aggregate UDFs
> - Tight integration with machine learning frameworks
>
> *Proposed API Changes*
> The following sketches some possibilities. I haven’t spent a lot of time
> thinking about the API (wrote it down in 5 mins) and I am not attached to
> this design at all. The main purpose of the SPIP is to get feedback on use
> cases and see how they can impact API design.
>
> Two things to consider are:
>
> 1. Python is dynamically typed, whereas DataFrames/SQL requires static,
> analysis time typing. This means users would need to specify the return type
> of their UDFs.
>
> 2. Ratio of input rows to output rows. We propose initially we require number
> of output rows to be the same as the number of input rows. In the future, we
> can consider relaxing this constraint with support for vectorized aggregate
> UDFs.
>
> Proposed API sketch (using examples):
>
> Use case 1. A function that defines all the columns of a DataFrame (similar
> to a “map” function):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_on_entire_df(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A Pandas data frame.
> """
> input[c] = input[a] + input[b]
> Input[d] = input[a] - input[b]
> return input
>
> spark.range(1000).selectExpr("id a", "id / 2 b")
> .mapBatches(my_func_on_entire_df)
> {code}
>
> Use case 2. A function that defines only one column (similar to existing
> UDFs):
>
> {code}
> @spark_udf(some way to describe the return schema)
> def my_func_that_returns_one_column(input):
> """ Some user-defined function.
>
> :param input: A Pandas DataFrame with two columns, a and b.
> :return: :class: A numpy array
> """
> return input[a] + input[b]
>
> my_func = udf(my_func_that_returns_one_column)
>
> df = spark.range(1000).selectExpr("id a", "id / 2 b")
> df.withColumn("c", my_func(df.a, df.b))
> {code}
>
>
>
> *Optional Design Sketch*
> I’m more concerned about getting proper feedback for API design. The
> implementation should be pretty straightforward and is not a huge concern at
> this point. We can leverage the same implementation for faster toPandas
> (using Arrow).
>
>
> *Optional Rejected Designs*
> See above.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
-
To unsubscribe, e-mail:
