[
https://issues.apache.org/jira/browse/FLUME-2973?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16497883#comment-16497883
]
Peter Turcsanyi commented on FLUME-2973:
----------------------------------------
The original patch ([^FLUME-2973.patch]) seems better to me, because in that
solution the order of acquiring / releasing the locks and the scope of the
locks are the same from both sides (both in {{HDFSEventSink}} and in
{{BucketWriter}}):
# acquiring lock on {{sfWriters}} map
# acquiring lock on {{BucketWriter}} instance
# releasing {{BucketWriter}} instance lock
# releasing {{sfWriters}} map lock
This was the original order in {{HDFSEventSink}} and the {{BucketWriter}}
implementation has been adjusted to it in [^FLUME-2973.patch].
The other patch ([^FLUME-2973-1.patch]) resolves the deadlock through
reordering the locking in {{BucketWriter}}, but in a different way as it was/is
in {{HDFSEventSink}}:
# acquiring lock on {{sfWriters}} map
# releasing {{sfWriters}} map lock
# acquiring lock on {{BucketWriter}} instance
# releasing {{BucketWriter}} instance lock
This solution prevents the deadlock too, but the different locking order might
lead to side effects / changed functionality.
For this reason, I would vote for [^FLUME-2973.patch].
However, it would be worth applying the test clean-up ({{BucketWriterBuilder}})
from [^FLUME-2973-1.patch].
> Deadlock in hdfs sink
> ---------------------
>
> Key: FLUME-2973
> URL: https://issues.apache.org/jira/browse/FLUME-2973
> Project: Flume
> Issue Type: Bug
> Components: Sinks+Sources
> Affects Versions: 1.7.0
> Reporter: Denes Arvay
> Assignee: Denes Arvay
> Priority: Critical
> Labels: hdfssink
> Fix For: 1.9.0
>
> Attachments: FLUME-2973-1.patch, FLUME-2973.patch
>
>
> Automatic close of BucketWriters (when open file count reached
> {{hdfs.maxOpenFiles}}) and the file rolling thread can end up in deadlock.
> When creating a new {{BucketWriter}} in {{HDFSEventSink}} it locks
> {{HDFSEventSink.sfWritersLock}} and the {{close()}} called in
> {{HDFSEventSink.sfWritersLock.removeEldestEntry}} tries to lock the
> {{BucketWriter}} instance.
> On the other hand if the file is being rolled in
> {{BucketWriter.close(boolean)}} it locks the {{BucketWriter}} instance first
> and in the close callback it tries to lock the {{sfWritersLock}}.
> The chances for this deadlock is higher when the {{hdfs.maxOpenFiles}}'s
> value is low (1).
> Script to reproduce:
> https://gist.github.com/adenes/96503a6e737f9604ab3ee9397a5809ff
> (put to
> {{flume-ng-sinks/flume-hdfs-sink/src/main/java/org/apache/flume/sink/hdfs}})
> Deadlock usually occurs before ~30 iterations.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
