[
https://issues.apache.org/jira/browse/SPARK-37089?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ankur Dave updated SPARK-37089:
-------------------------------
Description:
The task completion listener that closes the vectorized reader is registered
lazily in {Parquet,Orc}FileFormat#buildReaderWithPartitionValues(). Since task
completion listeners are executed in reverse order of registration, it always
runs before the Python writer thread can be interrupted.
This contradicts the assumption in
https://issues.apache.org/jira/browse/SPARK-37088 /
https://github.com/apache/spark/pull/34245 that task completion listeners are
registered bottom-up, preventing that fix from working properly.
was:
The task completion listener that closes the vectorized reader is registered
lazily in ParquetFileFormat#buildReaderWithPartitionValues(). Since task
completion listeners are executed in reverse order of registration, it always
runs before the Python writer thread can be interrupted.
This contradicts the assumption in
https://issues.apache.org/jira/browse/SPARK-37088 /
https://github.com/apache/spark/pull/34245 that task completion listeners are
registered bottom-up, preventing that fix from working properly.
> ParquetFileFormat/OrcFileFormat register task completion listeners lazily,
> causing Python writer thread to segfault when off-heap vectorized reader is
> enabled
> --------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: SPARK-37089
> URL: https://issues.apache.org/jira/browse/SPARK-37089
> Project: Spark
> Issue Type: Bug
> Components: PySpark, SQL
> Affects Versions: 3.0.3, 3.1.2, 3.2.0
> Reporter: Ankur Dave
> Assignee: Ankur Dave
> Priority: Major
>
> The task completion listener that closes the vectorized reader is registered
> lazily in {Parquet,Orc}FileFormat#buildReaderWithPartitionValues(). Since
> task completion listeners are executed in reverse order of registration, it
> always runs before the Python writer thread can be interrupted.
> This contradicts the assumption in
> https://issues.apache.org/jira/browse/SPARK-37088 /
> https://github.com/apache/spark/pull/34245 that task completion listeners are
> registered bottom-up, preventing that fix from working properly.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
