Jessica Cheng created LUCENE-4707:
-------------------------------------
Summary: Track file reference kept by readers that are opened
through the writer
Key: LUCENE-4707
URL: https://issues.apache.org/jira/browse/LUCENE-4707
Project: Lucene - Core
Issue Type: Bug
Components: core/index
Affects Versions: 4.0
Environment: Mac OS X 10.8.2 and Linux 2.6.32
Reporter: Jessica Cheng
We ran into a bug where files (mostly CFS) that are still referred to by our
NRT reader/searcher are deleted by IndexFileDeleter. As far as I can see from
the verbose logging and reading the code, it seems that the problem is the
creation and merging of these CFS files between hard commits. The files
referred to by hard commits are incRef’ed at commit checkpoints, so these files
won’t be deleted until they are decRef’ed when the commit is deleted according
to the DeletionPolicy (good). However, intermediate files that are created and
merged between the hard commits only have refs through the regular checkpoints,
so as soon as a new checkpoint no longer includes those files, they are
immediately deleted by the deleter. See the abridged verbose log lines that
illustrate this behavior:
IW 11 [Mon Jan 21 17:30:35 PST 2013; commitScheduler]: create compound file
_8.cfs
IFD 7 [Mon Jan 21 17:23:41 PST 2013; commitScheduler]: now checkpoint
"_0(4.0.0.2):C3_1(4.0.0.2):C7 _2(4.0.0.2):C16 _3(4.0.0.2):C21 _4(4.0.0.2):C5
_5(4.0.0.2):C5_6(4.0.0.2):C5 _7(4.0.0.2):C7 _8(4.0.0.2):c6" [9 segments ;
isCommit = false]
IFD 7 [Mon Jan 21 17:23:41 PST 2013; commitScheduler]: IncRef "_8.cfs":
pre-incr count is 0
IFD 7 [Mon Jan 21 17:23:42 PST 2013; commitScheduler]: now checkpoint
"_0(4.0.0.2):C3_1(4.0.0.2):C7 _2(4.0.0.2):C16 _3(4.0.0.2):C21 _4(4.0.0.2):C5
_5(4.0.0.2):C5 _6(4.0.0.2):C5 _7(4.0.0.2):C7 _8(4.0.0.2):c6 _9(4.0.0.2):c6" [10
segments ; isCommit = false]
IFD 7 [Mon Jan 21 17:23:42 PST 2013; commitScheduler]: IncRef "_8.cfs":
pre-incr count is 1
IFD 7 [Mon Jan 21 17:23:42 PST 2013; commitScheduler]: DecRef "_8.cfs":
pre-decr count is 2
IFD 7 [Mon Jan 21 17:23:42 PST 2013; Lucene Merge Thread #0]: now checkpoint
"_b(4.0.0.2):C81" [1 segments ; isCommit = false]
IFD 7 [Mon Jan 21 17:23:42 PST 2013; Lucene Merge Thread #0]: DecRef
"_8.cfs": pre-decr count is 1
IFD 7 [Mon Jan 21 17:23:42 PST 2013; Lucene Merge Thread #0]: delete "_8.cfs"
With this behavior, it seems no matter how frequently we refresh the reader
(unless we do it at every read), we’d run into the race where the reader still
holds a reference to the file that’s just been deleted by the deleter. My
proposal is to count the file reference handed out to the NRT reader/searcher
when writer.getReader(boolean) is called and decRef the files only when the
said reader is closed.
Please take a look and evaluate if my observations are correct and if the
proposal makes sense. Thanks!
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
