[
https://issues.apache.org/jira/browse/FLINK-7499?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16261075#comment-16261075
]
ASF GitHub Bot commented on FLINK-7499:
---------------------------------------
Github user NicoK commented on a diff in the pull request:
https://github.com/apache/flink/pull/4581#discussion_r152342364
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/io/disk/iomanager/AsynchronousBufferFileWriter.java
---
@@ -31,9 +31,26 @@ protected AsynchronousBufferFileWriter(ID channelID,
RequestQueue<WriteRequest>
super(channelID, requestQueue, CALLBACK, true);
}
+ /**
+ * Writes the given block asynchronously.
+ *
+ * @param buffer
+ * the buffer to be written (will be recycled when done)
--- End diff --
good catch, but actually, `SpillableSubpartition` doesn't do any recycling
itself: in its `finish()` method, it relies on the buffer being on-heap and
then garbage-collected, for the `add()` function, it relies on the caller, i.e.
`ResultPartition#add()` (which I also forgot to adapt).
> double buffer release in SpillableSubpartitionView
> --------------------------------------------------
>
> Key: FLINK-7499
> URL: https://issues.apache.org/jira/browse/FLINK-7499
> Project: Flink
> Issue Type: Sub-task
> Components: Network
> Affects Versions: 1.2.0, 1.1.4, 1.3.0, 1.1.5, 1.2.1, 1.2.2, 1.3.1, 1.4.0,
> 1.3.2, 1.3.3
> Reporter: Nico Kruber
> Assignee: Nico Kruber
> Priority: Blocker
>
> {{SpillableSubpartitionView#releaseMemory()}} recycles its buffers twice:
> once asynchronously after the write operation and once in
> {{SpillableSubpartitionView#releaseMemory()}} after adding the write
> operation to the queue.
> 1) if {{SpillableSubpartitionView#releaseMemory()}} hits first and the buffer
> is recycled, the memory region may already be reused despite the pending write
> 2) If, for some reason (probably only in tests like
> {{SpillableSubpartitionTest#testConsumeSpillablePartitionSpilledDuringConsume()}}?),
> the buffer is retained and to be used in parallel somewhere else it may also
> not be available anymore or contain corrupt data.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
