L. C. Hsieh created SPARK-60113:
-----------------------------------
Summary: Wait for the pipelined Python UDF writer to exit before
task completion cleanup
Key: SPARK-60113
URL: https://issues.apache.org/jira/browse/SPARK-60113
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 5.0.0
Reporter: L. C. Hsieh
In pipelined Python UDF mode (SPARK-56642), the task completion listener in
BasePythonRunner calls writerFuture.cancel(true) and then writerFuture.get() to
wait for the writer thread to exit before later listeners free the input rows'
memory (SPARK-33277). After a successful cancel(true), FutureTask.get() throws
CancellationException immediately and does not wait for the runnable to finish.
The writer can therefore still be inside HybridRowQueue.add() (lock-free mode
writes directly into the current page) when the queue-close listener frees that
page, which is a use-after-free and can crash the JVM for off-heap pages. A
createNewQueue racing with the close can also leak its page until task cleanup.
The listener should wait for the writer to actually exit, e.g. on a latch the
writer counts down in a finally block.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
