[
https://issues.apache.org/jira/browse/LUCENE-1282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12596964#action_12596964
]
Paul Smith commented on LUCENE-1282:
------------------------------------
Throwing up an idea here for consideration. I'm sure it could be shot down,
but I thought I'd raise it just in case it hasn't already been considered and
discarded..
One of the _classic_ problems between -client and -server mode is the way the
CPU registers are used. Is it possible that some of the fields are suffering
from concurrency issues? I was wondering if, say, BufferedInfexOutput.buffer*
may need to be marked volatile ?
One easy way to test if this makes a difference is to just try switching
between explicit use of '-client' and '-server'. Most newer machines (even
desktops & laptops) appear to qualify for Sun's 'am I a server-class machine'
check. By switching to -client, if these problems disappear, this to me would
smell more and more like a 'volatile' like behaviour, because AIUI, -server
will be more aggressive with some of it's register optimizations and I've seen
behaviour just like this where variables that have clearly been written, the
changes are not 'appearing' on the other side. Even the same thread marking
the change can be switched across to a different CPU right in the middle, and
could see different results.
Of course those people with lots of concurrency experience can probably dismiss
this theory in a second, but that's fine.
> Sun hotspot compiler bug in 1.6.0_04/05 affects Lucene
> ------------------------------------------------------
>
> Key: LUCENE-1282
> URL: https://issues.apache.org/jira/browse/LUCENE-1282
> Project: Lucene - Java
> Issue Type: Bug
> Components: Index
> Affects Versions: 2.3, 2.3.1
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Priority: Minor
> Fix For: 2.4
>
> Attachments: corrupt_merge_out15.txt
>
>
> This is not a Lucene bug. It's an as-yet not fully characterized Sun
> JRE bug, as best I can tell. I'm opening this to gather all things we
> know, and to work around it in Lucene if possible, and maybe open an
> issue with Sun if we can reduce it to a compact test case.
> It's hit at least 3 users:
>
> http://mail-archives.apache.org/mod_mbox/lucene-java-user/200803.mbox/[EMAIL
> PROTECTED]
>
> http://mail-archives.apache.org/mod_mbox/lucene-solr-user/200804.mbox/[EMAIL
> PROTECTED]
>
> http://mail-archives.apache.org/mod_mbox/lucene-java-user/200805.mbox/[EMAIL
> PROTECTED]
> It's specific to at least JRE 1.6.0_04 and 1.6.0_05, that affects
> Lucene. Whereas 1.6.0_03 works OK and it's unknown whether 1.6.0_06
> shows it.
> The bug affects bulk merging of stored fields. When it strikes, the
> segment produced by a merge is corrupt because its fdx file (stored
> fields index file) is missing one document. After iterating many
> times with the first user that hit this, adding diagnostics &
> assertions, its seems that a call to fieldsWriter.addDocument some
> either fails to run entirely, or, fails to invoke its call to
> indexStream.writeLong. It's as if when hotspot compiles a method,
> there's some sort of race condition in cutting over to the compiled
> code whereby a single method call fails to be invoked (speculation).
> Unfortunately, this corruption is silent when it occurs and only later
> detected when a merge tries to merge the bad segment, or an
> IndexReader tries to open it. Here's a typical merge exception:
> {code}
> Exception in thread "Thread-10"
> org.apache.lucene.index.MergePolicy$MergeException:
> org.apache.lucene.index.CorruptIndexException:
> doc counts differ for segment _3gh: fieldsReader shows 15999 but
> segmentInfo shows 16000
> at
> org.apache.lucene.index.ConcurrentMergeScheduler$MergeThread.run(ConcurrentMergeScheduler.java:271)
> Caused by: org.apache.lucene.index.CorruptIndexException: doc counts differ
> for segment _3gh: fieldsReader shows 15999 but segmentInfo shows 16000
> at
> org.apache.lucene.index.SegmentReader.initialize(SegmentReader.java:313)
> at org.apache.lucene.index.SegmentReader.get(SegmentReader.java:262)
> at org.apache.lucene.index.SegmentReader.get(SegmentReader.java:221)
> at
> org.apache.lucene.index.IndexWriter.mergeMiddle(IndexWriter.java:3099)
> at org.apache.lucene.index.IndexWriter.merge(IndexWriter.java:2834)
> at
> org.apache.lucene.index.ConcurrentMergeScheduler$MergeThread.run(ConcurrentMergeScheduler.java:240)
> {code}
> and here's a typical exception hit when opening a searcher:
> {code}
> org.apache.lucene.index.CorruptIndexException: doc counts differ for segment
> _kk: fieldsReader shows 72670 but segmentInfo shows 72671
> at
> org.apache.lucene.index.SegmentReader.initialize(SegmentReader.java:313)
> at org.apache.lucene.index.SegmentReader.get(SegmentReader.java:262)
> at org.apache.lucene.index.SegmentReader.get(SegmentReader.java:230)
> at
> org.apache.lucene.index.DirectoryIndexReader$1.doBody(DirectoryIndexReader.java:73)
> at
> org.apache.lucene.index.SegmentInfos$FindSegmentsFile.run(SegmentInfos.java:636)
> at
> org.apache.lucene.index.DirectoryIndexReader.open(DirectoryIndexReader.java:63)
> at org.apache.lucene.index.IndexReader.open(IndexReader.java:209)
> at org.apache.lucene.index.IndexReader.open(IndexReader.java:173)
> at
> org.apache.lucene.search.IndexSearcher.<init>(IndexSearcher.java:48)
> {code}
> Sometimes, adding -Xbatch (forces up front compilation) or -Xint
> (disables compilation) to the java command line works around the
> issue.
> Here are some of the OS's we've seen the failure on:
> {code}
> SuSE 10.0
> Linux phoebe 2.6.13-15-smp #1 SMP Tue Sep 13 14:56:15 UTC 2005 x86_64
> x86_64 x86_64 GNU/Linux
> SuSE 8.2
> Linux phobos 2.4.20-64GB-SMP #1 SMP Mon Mar 17 17:56:03 UTC 2003 i686
> unknown unknown GNU/Linux
> Red Hat Enterprise Linux Server release 5.1 (Tikanga)
> Linux lab8.betech.virginia.edu 2.6.18-53.1.14.el5 #1 SMP Tue Feb 19
> 07:18:21 EST 2008 i686 i686 i386 GNU/Linux
> {code}
> I've already added assertions to Lucene to detect when this bug
> strikes, but since assertions are not usually enabled, I plan to add a
> real check to catch when this bug strikes *before* we commit the merge
> to the index. This way we can detect & quarantine the failure and
> prevent corruption from entering the index.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
