[ 
https://issues.apache.org/jira/browse/SPARK-17460?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Chris Perluss updated SPARK-17460:
----------------------------------
    Component/s: SQL

> Dataset.joinWith broadcasts gigabyte sized table, causes OOM Exception
> ----------------------------------------------------------------------
>
>                 Key: SPARK-17460
>                 URL: https://issues.apache.org/jira/browse/SPARK-17460
>             Project: Spark
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions: 2.0.0
>         Environment: Spark 2.0 in local mode as well as on GoogleDataproc
>            Reporter: Chris Perluss
>
> Dataset.joinWith is performing a BroadcastJoin on a table that is gigabytes 
> in size due to the dataset.logicalPlan.statistics.sizeInBytes < 0.
> The issue is that org.apache.spark.sql.types.ArrayType.defaultSize is of 
> datatype Int.  In my dataset, there is an Array column whose data size 
> exceeds the limits of an Int and so the data size becomes negative.
> The issue can be repeated by running this code in REPL:
> val ds = (0 to 10000).map( i => (i, Seq((i, Seq((i, "This is really not that 
> long of a string")))))).toDS()
> // You might have to remove private[sql] from Dataset.logicalPlan to get this 
> to work
> val stats = ds.logicalPlan.statistics
> yields
> stats: org.apache.spark.sql.catalyst.plans.logical.Statistics = 
> Statistics(-1890686892,false)
> This causes joinWith to performWith to perform a broadcast join even tho my 
> data is gigabytes in size, which of course causes the executors to run out of 
> memory.
> Setting spark.sql.autoBroadcastJoinThreshold=-1 does not help because the 
> logicalPlan.statistics.sizeInBytes is a large negative number and thus it is 
> less than the join threshold of -1.
> I've been able to work around this issue by setting 
> autoBroadcastJoinThreshold to a very large negative number.



--
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

Reply via email to