[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15963367#comment-15963367
]
Charles Pritchard commented on SPARK-19352:
---
Does this fix the issue in SPARK-18934 ?
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15963365#comment-15963365
]
Charles Pritchard commented on SPARK-19352:
---
[~cloud_fan] Yes, Hive relies on sorting optimizations for running map side
joins. DISTRIBUTE BY and SORT BY can be used to manually output data into
single sorted files per partition.
Hive will ensure sorting when running INSERT OVERWRITE statements, when a table
is created with PARTITIONED BY... CLUSTERED BY... SORTED BY ... INTO 1 BUCKETS.
Spark also reads the Hive metastore to detect when files are already sorted,
and runs optimizations.
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15961686#comment-15961686
]
Wenchen Fan commented on SPARK-19352:
-
I don't think Spark will provide API support for this feature(Does hive really
have?), but the implementation is quite stable now, so you can follow the
example in this ticket to write out sorted data.
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15961389#comment-15961389
]
Charles Pritchard commented on SPARK-19352:
---
[~cloud_fan] Is there something on the roadmap to get that guarantee? We need
guaranteed sorting from a general performance perspective, but it's also a
baseline feature of Hive (AKA: "SORT BY") to be able to sort data into a file
in a partition.
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15884032#comment-15884032
]
Wenchen Fan commented on SPARK-19352:
-
I'm going to mark it as `not a problem`. Spark doesn't guarantee the sorting
when writing data out, although now the data will sorted as you expected in
your example, but it depends on the implementation details and may change in
the future.
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15884016#comment-15884016
]
Liang-Chi Hsieh commented on SPARK-19352:
-
I think this is in fact solved by SPARK-19563. [~cloud_fan]
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15851387#comment-15851387
]
Wenchen Fan commented on SPARK-19352:
-
DataFrameWriter doesn't allow users to write data out orderly, so unfortunately
you can't do this. This makes sense because sorted data files are not useful:
when you read it back, Spark can't get the ordering information back and
optimize for it.
However, in your example, the data files should be sorted, according to how
Spark write partitioned data. If the input data is already partitioned, the
writer doesn't need to sort the data by partition columns anymore.
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15843736#comment-15843736
]
Apache Spark commented on SPARK-19352:
--
User 'viirya' has created a pull request for this issue:
https://github.com/apache/spark/pull/16724
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
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
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15836653#comment-15836653
]
Ivan Gozali commented on SPARK-19352:
-
Does this mean that {{Dataset.write.partitionBy()}} performs a
repartition/shuffling of the dataset even if it's already partitioned by the
same column beforehand? If this is indeed the case, then with the current API,
I haven't seen a way to do what I'd like to do, which is emit one CSV sorted by
timestamp per user ID (or at least, multiple sorted CSVs but can easily be
merged by concatenating them). This seems like it should have been a very
simple thing to do in Spark.
DataFrameWriter.sortBy() looked promising, but when I used it, it gave me the
following error:
{code}
org.apache.spark.sql.AnalysisException: 'save' does not support bucketing right
now;
at
org.apache.spark.sql.DataFrameWriter.assertNotBucketed(DataFrameWriter.scala:314)
at org.apache.spark.sql.DataFrameWriter.save(DataFrameWriter.scala:207)
at org.apache.spark.sql.DataFrameWriter.save(DataFrameWriter.scala:198)
{code}
Is it not possible to do this?
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
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
[jira] [Commented] (SPARK-19352) Sorting issues on relatively big datasets
[
https://issues.apache.org/jira/browse/SPARK-19352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15836626#comment-15836626
]
Sean Owen commented on SPARK-19352:
---
You repartition by userID after sorting -- is that not probably the issue?
> Sorting issues on relatively big datasets
> -
>
> Key: SPARK-19352
> URL: https://issues.apache.org/jira/browse/SPARK-19352
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
>Affects Versions: 2.1.0
> Environment: Spark version 2.1.0
> Using Scala version 2.11.8, Java HotSpot(TM) 64-Bit Server VM, 1.8.0_102
> macOS 10.12.3
>Reporter: Ivan Gozali
>
> _More details, including the script to generate the synthetic dataset
> (requires pandas and numpy) are in this GitHub gist._
> https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f
> Given a relatively large synthetic time series dataset of various users
> (4.1GB), when attempting to:
> * partition this dataset by user ID
> * sort the time series data for each user by timestamp
> * write each partition to a single CSV file
> then some files are unsorted in a very specific manner. In one of the
> supposedly sorted files, the rows looked as follows:
> {code}
> 2014-01-01T00:00:00.000-08:00,-0.07,0.39,-0.39
> 2014-12-31T02:07:30.000-08:00,0.34,-0.62,-0.22
> 2014-01-01T00:00:05.000-08:00,-0.07,-0.52,0.47
> 2014-12-31T02:07:35.000-08:00,-0.15,-0.13,-0.14
> 2014-01-01T00:00:10.000-08:00,-1.31,-1.17,2.24
> 2014-12-31T02:07:40.000-08:00,-1.28,0.88,-0.43
> {code}
> The above is attempted using the following Scala/Spark code:
> {code}
> val inpth = "/tmp/gen_data_3cols_small"
> spark
> .read
> .option("inferSchema", "true")
> .option("header", "true")
> .csv(inpth)
> .repartition($"userId")
> .sortWithinPartitions("timestamp")
> .write
> .partitionBy("userId")
> .option("header", "true")
> .csv(inpth + "_sorted")
> {code}
> This issue is not seen when using a smaller sized dataset by making the time
> span smaller (354MB, with the same number of columns).
--
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
