[
https://issues.apache.org/jira/browse/LUCENE-3551?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Simon Willnauer updated LUCENE-3551:
------------------------------------
Attachment: LUCENE-3551.patch
I isolated the problem in SOLR-2861 into a lucene testcase that fails
reproduceable & very very quickly. (passed on 3.x too though)
with the fixes in this patch it passes reliably. The problem among the others I
already explained is that once DW has flushed all threads and put all deletes
in the frozenPacketBuffer we have a little window where those changes are not
taken into account. This only happens if I only flush deletes (no documents)
since we prune the frozenBufferedDeletes before we checkpoint so nrtIsCurrent
doesn't see those changes for a little while.
> Yet another race in IW#nrtIsCurrent
> -----------------------------------
>
> Key: LUCENE-3551
> URL: https://issues.apache.org/jira/browse/LUCENE-3551
> Project: Lucene - Java
> Issue Type: Bug
> Components: core/index
> Affects Versions: 4.0
> Reporter: Simon Willnauer
> Fix For: 4.0
>
> Attachments: LUCENE-3551.patch, LUCENE-3551.patch, LUCENE-3551.patch,
> LUCENE-3551.patch
>
>
> In IW#nrtIsCurrent looks like this:
> {code}
> synchronized boolean nrtIsCurrent(SegmentInfos infos) {
> ensureOpen();
> return infos.version == segmentInfos.version && !docWriter.anyChanges()
> && !bufferedDeletesStream.any();
> }
> {code}
> * the version changes once we checkpoint the IW
> * docWriter has changes if there are any docs in ram or any deletes in the
> delQueue
> * bufferedDeletes contain all frozen del packages from the delQueue
> yet, what happens is 1. we decrement the numDocsInRam in DWPT#doAfterFlush
> (which is executed during DWPT#flush) but before we checkpoint. 2. if we
> freeze deletes (empty the delQueue) we put them in the flushQueue to maintain
> the order. This means they are not yet in the bufferedDeleteStream.
> Bottom line, there is a window where we could see IW#nrtIsCurrent returning
> true if we check within this particular window. Phew, I am not 100% sure if
> that is the reason for our latest failure in SOLR-2861 but from what the logs
> look like this could be what happens. If we randomly hit low values for
> maxBufferedDocs & maxBufferedDeleteTerms this is absolutely possible.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
