[ 
https://issues.apache.org/jira/browse/SPARK-61765?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

L. C. Hsieh updated SPARK-61765:
--------------------------------
    Affects Version/s: 4.4.0
                           (was: 5.0.0)

> Report the Python worker idle timeout when the read from the killed worker 
> fails
> --------------------------------------------------------------------------------
>
>                 Key: SPARK-61765
>                 URL: https://issues.apache.org/jira/browse/SPARK-61765
>             Project: Spark
>          Issue Type: Bug
>          Components: PySpark
>    Affects Versions: 4.4.0
>            Reporter: L. C. Hsieh
>            Priority: Major
>
> After the idle timeout kills a Python worker 
> (spark.sql.execution.pyspark.udf.killOnIdleTimeout), the read from the worker 
> does not always return the end of the stream. The connection of a killed 
> worker can be reset, for example when the worker had unread input, and the 
> read then throws an IOException (Connection reset).
> IdleTimeoutHandler.throwIfKilled only looks at the value a read returned, so 
> in that case the task fails with "Python worker exited unexpectedly 
> (crashed)", together with the hint to enable the fault handler, instead of 
> "Python worker process terminated due to idle timeout (timeout: N seconds)". 
> This affects the three streams that use the handler: 
> BasePythonRunner.ReaderInputStream, the pipelined socket stream, and 
> PythonPlannerRunner.WorkerInputStream.
> A read (or a write on the reading thread) that fails after the handler killed 
> the worker should be reported as the idle timeout, with the I/O failure as 
> the cause. A failure with no kill should be reported as today.
> This was pointed out in the review of SPARK-61720.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to