lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Itamar Syn-Hershko <ita...@code972.com>
Subject Re: Corrupt index
Date Thu, 14 Jun 2012 17:41:15 GMT
I'm quite certain this shouldn't happen also when Commit wasn't called.

Mike, can you comment on that?

On Thu, Jun 14, 2012 at 8:03 PM, Christopher Currens <
currens.chris@gmail.com> wrote:

> Well, the only thing I see is that there is no place where writer.Commit()
> is called in the delegate assigned to corpusReader.OnDocument.  I know that
> lucene is very transactional, and at least in 3.x, the writer will never
> auto commit to the index.  You can write millions of documents, but if
> commit is never called, those documents aren't actually part of the index.
>  Committing isn't a cheap operation, so you definitely don't want to do it
> on every document.
>
> You can test it yourself with this (naive) solution.  Right below the
> writer.SetUseCompoundFile(false) line, add "int numDocsAdded = 0;".  At the
> end of the corpusReader.OnDocument delegate add:
>
> // Example only.  I wouldn't suggest committing this often
> if(++numDocsAdded % 5 == 0)
> {
>    writer.Commit();
> }
>
> I had the application crash for real on this file:
>
> http://dumps.wikimedia.org/gawiktionary/20120613/gawiktionary-20120613-pages-meta-history.xml.bz2
> ,
> about 20% into the operation.  Without the commit, the index is empty.  Add
> it in, and I get 755 files in the index after it crashes.
>
>
> Thanks,
> Christopher
>
> On Wed, Jun 13, 2012 at 6:13 PM, Itamar Syn-Hershko <itamar@code972.com
> >wrote:
>
> > Yes, reproduced in first try. See attached program - I referenced it to
> > current trunk.
> >
> >
> > On Thu, Jun 14, 2012 at 3:54 AM, Itamar Syn-Hershko <itamar@code972.com
> >wrote:
> >
> >> Christopher,
> >>
> >> I used the IndexBuilder app from here
> >> https://github.com/synhershko/Talks/tree/master/LuceneNeatThings with a
> >> 8.5GB wikipedia dump.
> >>
> >> After running for 2.5 days I had to forcefully close it (infinite loop
> in
> >> the wiki-markdown parser at 92%, go figure), and the 40-something GB
> index
> >> I had by then was unusable. I then was able to reproduce this
> >>
> >> Please note I now added a few safe-guards you might want to remove to
> >> make sure the app really crashes on process kill.
> >>
> >> I'll try to come up with a better way to reproduce this - hopefully Mike
> >> will be able to suggest better ways than manual process kill...
> >>
> >> On Thu, Jun 14, 2012 at 1:41 AM, Christopher Currens <
> >> currens.chris@gmail.com> wrote:
> >>
> >>> Mike, The codebase for lucene.net should be almost identical to java's
> >>> 3.0.3 release, and LUCENE-1044 is included in that.
> >>>
> >>> Itamar, are you committing the index regularly?  I only ask because I
> >>> can't
> >>> reproduce it myself by forcibly terminating the process while it's
> >>> indexing.  I've tried both 3.0.3 and 2.9.4.  If I don't commit at all
> and
> >>> terminate the process (even with a 10,000 4K documents created), there
> >>> will
> >>> be no documents in the index when I open it in luke, which I expect.
>  If
> >>> I
> >>> commit at 10,000 documents, and terminate it a few thousand after that,
> >>> the
> >>> index has the first ten thousand that were committed.  I've even
> >>> terminated
> >>> it *while* a second commit was taking place, and it still had all of
> the
> >>> documents I expected.
> >>>
> >>> It may be that I'm not trying to reproducing it correctly.  Do you
> have a
> >>> minimal amount of code that can reproduce it?
> >>>
> >>>
> >>> Thanks,
> >>> Christopher
> >>>
> >>> On Wed, Jun 13, 2012 at 9:31 AM, Michael McCandless <
> >>> lucene@mikemccandless.com> wrote:
> >>>
> >>> > Hi Itamar,
> >>> >
> >>> > One quick question: does Lucene.Net include the fixes done for
> >>> > LUCENE-1044 (to fsync files on commit)?  Those are very important for
> >>> > an index to be intact after OS/JVM crash or power loss.
> >>> >
> >>> > More responses below:
> >>> >
> >>> > On Tue, Jun 12, 2012 at 8:20 PM, Itamar Syn-Hershko <
> >>> itamar@code972.com>
> >>> > wrote:
> >>> >
> >>> > > I'm a Lucene.Net committer, and there is a chance we have a bug
in
> >>> our
> >>> > > FSDirectory implementation that causes indexes to get corrupted
> when
> >>> > > indexing is cut while the IW is still open. As it roots from some
> >>> > > retroactive fixes you made, I'd appreciate your feedback.
> >>> > >
> >>> > > Correct me if I'm wrong, but by design Lucene should be able to
> >>> recover
> >>> > > rather quickly from power failures or app crashes. Since existing
> >>> segment
> >>> > > files are read only, only new segments that are still being written
> >>> can
> >>> > get
> >>> > > corrupted. Hence, recovering from worst-case scenarios is done
by
> >>> simply
> >>> > > removing the write.lock file. The worst that could happen then
is
> >>> having
> >>> > the
> >>> > > last segment damaged, and that can be fixed by removing those
> files,
> >>> > > possibly by running CheckIndex on the index.
> >>> >
> >>> > You shouldn't even have to run CheckIndex ... because (as of
> >>> > LUCENE-1044) we now fsync all segment files before writing the new
> >>> > segments_N file, and then removing old segments_N files (and any
> >>> > segments that are no longer referenced).
> >>> >
> >>> > You do have to remove the write.lock if you aren't using
> >>> > NativeFSLockFactory (but this has been the default lock impl for a
> >>> > while now).
> >>> >
> >>> > > Last week I have been playing with rather large indexes and crashed
> >>> my
> >>> > app
> >>> > > while it was indexing. I wasn't able to open the index, and Luke
> was
> >>> even
> >>> > > kind enough to wipe the index folder clean even though I opened
it
> in
> >>> > > read-only mode. I re-ran this, and after another crash running
> >>> CheckIndex
> >>> > > revealed nothing - the index was detected to be an empty one.
I am
> >>> not
> >>> > > entirely sure what could be the cause for this, but I suspect
it
> has
> >>> > > been corrupted by the crash.
> >>> >
> >>> > Had no commit completed (no segments file written)?
> >>> >
> >>> > If you don't fsync then all sorts of crazy things are possible...
> >>> >
> >>> > > I've been looking at these:
> >>> > >
> >>> > >
> >>> >
> >>>
> https://issues.apache.org/jira/browse/LUCENE-3418?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
> >>> > >
> >>> >
> >>>
> https://issues.apache.org/jira/browse/LUCENE-2328?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
> >>> >
> >>> > (And LUCENE-1044 before that ... it was LUCENE-1044 that LUCENE-2328
> >>> > broke...).
> >>> >
> >>> > > And it seems like this is what I was experiencing. Mike and Mark
> will
> >>> > > probably be able to tell if this is what they saw or not, but
as
> far
> >>> as I
> >>> > > can tell this is not an expected behavior of a Lucene index.
> >>> >
> >>> > Definitely not expected behavior: assuming nothing is flipping bits,
> >>> > then on OS/JVM crash or power loss your index should be fine, just
> >>> > reverted to the last successful commit.
> >>> >
> >>> > > What I'm looking for at the moment is some advice on what
> FSDirectory
> >>> > > implementation to use to make sure no corruption can happen. The
> 3.4
> >>> > version
> >>> > > (which is where LUCENE-3418 was committed to) seems to handle
a
> lot
> >>> of
> >>> > > things the 3.0 doesn't, but on the other hand LUCENE-3418 was
> >>> introduced
> >>> > by
> >>> > > changes made to the 3.0 codebase.
> >>> >
> >>> > Hopefully it's just that you are missing fsync!
> >>> >
> >>> > > Also, is there any test in the suite checking for those scenarios?
> >>> >
> >>> > Our test framework has a sneaky MockDirectoryWrapper that, after a
> >>> > test finishes, goes and corrupts any unsync'd files and then verifies
> >>> > the index is still OK... it's good because it'll catch any times we
> >>> > are missing calls t sync, but, it's not low level enough such that
if
> >>> > FSDir is failing to actually call fsync (that wsa the bug in
> >>> > LUCENE-3418) then it won't catch that...
> >>> >
> >>> > Mike McCandless
> >>> >
> >>> > http://blog.mikemccandless.com
> >>> >
> >>>
> >>
> >>
> >
>

Mime
View raw message