On Tue, Apr 21, 2009 at 6:40 PM, Christiaan Fluit <christiaan.fl...@aduna-software.com> wrote:
> I may be on to something already. > > I just looked at the commitMerge code and was surprised to see that the > commitMerge message that is almost at the beginning wasn't printed. Then I > saw the "if (hitOOM) return false;" part that takes place before that. I > think that this can only mean that an OOME was encountered at some point in > time. Very interesting! I like this theory... > Now, the fact is that in my indexing code I do a catch(Throwable) in several > places. I do this particularly because JET handles OOMEs in a very, very > nasty way. Often you will just get an error dialog and then it quits the > entire application. Therefore, my client code catches, logs and swallows the > OOME before the JET runtime can intercept it. *Usually*, the application can > then recover gracefully and continue processing the rest of the information. > > Catching a OOME that results from the operation of a text extraction library > is one thing (and a fact of life really), but perhaps there are also OOME's > that occur during Lucene processing. > > I remember seeing those in the past with the original Java code, when very > large Strings were being tokenized and I got an OOME with a deep Lucene > stacktrace. I copied one such stacktrace that I have saved at the end of > this mail. > > I see some caught and swallowed OOME's in my log file but unfortunately they > are without a stacktrace - probably again a JET issue. I can run the normal > Java build though to see if such OOMEs occur on this dataset. > > Now, I wonder: > > - when the IW is in auto-commit mode, can the failed processing of a > Document due to an OOME have an impact on the processing of subsequent > Documents or the merge/optimize operations? Can the index(writer) become > corrupt and result in problems such as these? On hitting OOME, Lucene refuses to commit any further changes to the index. Ie, you must at that point abandon the writer (writer.rollback()). We do this as a defense against the possibility that an OOME might otherwise cause index corruption. > - even though the commitMerge returns false, it should probably not get into > an infinite loop. Is this an internal Lucene problem or is there something I > can/should do about it myself? Yes, something is wrong with Lucene's handling of OOME. It certainly should not lead to infinite merge attempts. I'll dig (once back from vacation) to see if I can find this path. Likely we need to prevent launching of new merges after an OOME. I think you must've happened to hit OOME when a merge was running. > - more generally, what is the recommended behavior when I get an OOME during > Lucene processing, particularly IW.addDocument? Should the IW be able to > recover by itself or is there some sort of rollback I need to perform? You need to call writer.rollback(). > Again, note that my index is in auto-commit mode (though I had hoped to let > go of that too, it's only for historic reasons). I think being in autoCommit mode shouldn't affect this, ie likely you'd hit the infinite loop with autoCommit=false too. Mike --------------------------------------------------------------------- To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org For additional commands, e-mail: java-user-h...@lucene.apache.org