[jira] [Commented] (SPARK-23442) Reading from partitioned and bucketed table uses only bucketSpec.numBuckets partitions in all cases
[
https://issues.apache.org/jira/browse/SPARK-23442?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16494806#comment-16494806
]
Apache Spark commented on SPARK-23442:
--
User 'wangyum' has created a pull request for this issue:
https://github.com/apache/spark/pull/21460
> Reading from partitioned and bucketed table uses only bucketSpec.numBuckets
> partitions in all cases
> ---
>
> Key: SPARK-23442
> URL: https://issues.apache.org/jira/browse/SPARK-23442
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL
>Affects Versions: 2.2.1
>Reporter: Pranav Rao
>Priority: Major
>
> Through the DataFrameWriter[T] interface I have created a external HIVE table
> with 5000 (horizontal) partitions and 50 buckets in each partition. Overall
> the dataset is 600GB and the provider is Parquet.
> Now this works great when joining with a similarly bucketed dataset - it's
> able to avoid a shuffle.
> But any action on this Dataframe(from _spark.table("tablename")_), works with
> only 50 RDD partitions. This is happening because of
> [createBucketedReadRDD|https://github.com/apachttps:/github.com/apache/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.she/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.sc].
> So the 600GB dataset is only read through 50 tasks, which makes this
> partitioning + bucketing scheme not useful.
> I cannot expose the base directory of the parquet folder for reading the
> dataset, because the partition locations don't follow a (basePath + partSpec)
> format.
> Meanwhile, are there workarounds to use higher parallelism while reading such
> a table?
> Let me know if I can help in any way.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-23442) Reading from partitioned and bucketed table uses only bucketSpec.numBuckets partitions in all cases
[
https://issues.apache.org/jira/browse/SPARK-23442?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16368117#comment-16368117
]
Pranav Rao commented on SPARK-23442:
Repartitioning is unlikely to be helpful to a user because:
* The map part of repartition is still limited to num_buckets, so it's going to
be very slow and not utilise available parallelism.
* The user would have pre-partitioned and bucketed his dataset and persisted it
so, purely to avoid repartitioning/shuffle at read time. So the purpose of this
feature is lost.
> Reading from partitioned and bucketed table uses only bucketSpec.numBuckets
> partitions in all cases
> ---
>
> Key: SPARK-23442
> URL: https://issues.apache.org/jira/browse/SPARK-23442
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL
>Affects Versions: 2.2.1
>Reporter: Pranav Rao
>Priority: Major
>
> Through the DataFrameWriter[T] interface I have created a external HIVE table
> with 5000 (horizontal) partitions and 50 buckets in each partition. Overall
> the dataset is 600GB and the provider is Parquet.
> Now this works great when joining with a similarly bucketed dataset - it's
> able to avoid a shuffle.
> But any action on this Dataframe(from _spark.table("tablename")_), works with
> only 50 RDD partitions. This is happening because of
> [createBucketedReadRDD|https://github.com/apachttps:/github.com/apache/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.she/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.sc].
> So the 600GB dataset is only read through 50 tasks, which makes this
> partitioning + bucketing scheme not useful.
> I cannot expose the base directory of the parquet folder for reading the
> dataset, because the partition locations don't follow a (basePath + partSpec)
> format.
> Meanwhile, are there workarounds to use higher parallelism while reading such
> a table?
> Let me know if I can help in any way.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-23442) Reading from partitioned and bucketed table uses only bucketSpec.numBuckets partitions in all cases
[
https://issues.apache.org/jira/browse/SPARK-23442?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16366898#comment-16366898
]
Marco Gaido commented on SPARK-23442:
-
I am not sure it is what you are looking for, but you can repartition the
resulting DataFrame in order to have more partitions.
> Reading from partitioned and bucketed table uses only bucketSpec.numBuckets
> partitions in all cases
> ---
>
> Key: SPARK-23442
> URL: https://issues.apache.org/jira/browse/SPARK-23442
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL
>Affects Versions: 2.2.1
>Reporter: Pranav Rao
>Priority: Major
>
> Through the DataFrameWriter[T] interface I have created a external HIVE table
> with 5000 (horizontal) partitions and 50 buckets in each partition. Overall
> the dataset is 600GB and the provider is Parquet.
> Now this works great when joining with a similarly bucketed dataset - it's
> able to avoid a shuffle.
> But any action on this Dataframe(from _spark.table("tablename")_), works with
> only 50 RDD partitions. This is happening because of
> [createBucketedReadRDD|https://github.com/apachttps:/github.com/apache/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.she/spark/blob/branch-2.3/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.sc].
> So the 600GB dataset is only read through 50 tasks, which makes this
> partitioning + bucketing scheme not useful.
> I cannot expose the base directory of the parquet folder for reading the
> dataset, because the partition locations don't follow a (basePath + partSpec)
> format.
> Meanwhile, are there workarounds to use higher parallelism while reading such
> a table?
> Let me know if I can help in any way.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
