[
https://issues.apache.org/jira/browse/SPARK-17728?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15534923#comment-15534923
]
Jacob Eisinger commented on SPARK-17728:
----------------------------------------
Thanks for the explanation, but I still think this is an issue.
If Spark assumed their was no side effects and optimize accordingly, their
would be not issue: the UDF would be called once per row (1). However, Spark
calls a costly function many times leading to inefficiency.
In our production code, we have a function that takes in a long string and
classifies it under a number of different dimensions. This is a very CPU
intensive operation and is a pure function . Obviously, if Spark's optimizer
calls the functions multiple times, this is _not_ optimal in this scenario.
I think it is intuitive to most that the following code would call the UDF once
per row (1):
{code}
val exploded = as
.withColumn("structured_information", fUdf('a))
.withColumn("plus_one", 'structured_information("plusOne"))
.withColumn("squared", 'structured_information("squared"))
{code}
However, Spark calls the UDF three times per row! Is this what you would
expect? What am I missing?
(1) - "Once per row" - except when the row needs to recomputed such as when
workers are lost.
(2) - I attempted to model the long operation via Thread.sleep(); as you
mentioned this does have a slight side effect. Maybe I should have summed the
first billion counting numbers to illustrate the slow down?
> UDFs are run too many times
> ---------------------------
>
> Key: SPARK-17728
> URL: https://issues.apache.org/jira/browse/SPARK-17728
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 2.0.0
> Environment: Databricks Cloud / Spark 2.0.0
> Reporter: Jacob Eisinger
> Priority: Minor
> Attachments: over_optimized_udf.html
>
>
> h3. Background
> Llonger running processes that might run analytics or contact external
> services from UDFs. The response might not just be a field, but instead a
> structure of information. When attempting to break out this information, it
> is critical that query is optimized correctly.
> h3. Steps to Reproduce
> # Create some sample data.
> # Create a UDF that returns a multiple attributes.
> # Run UDF over some data.
> # Create new columns from the multiple attributes.
> # Observe run time.
> h3. Actual Results
> The UDF is executed *multiple times* _per row._
> h3. Expected Results
> The UDF should only be executed *once* _per row._
> h3. Workaround
> Cache the Dataset after UDF execution.
> h3. Details
> For code and more details, see [^over_optimized_udf.html]
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
