Well I've done two or three more runs with Sun java build 1.6.0_10-rc-b28, with and without the infoStream change, and I can't reproduce the problem. So maybe everything was indeed the fault of the release version of Sun hotspot. Not too confident about that. Maybe instead I let the disk get too full before or something.
In any case I'm unfortunately not sure I have time to keep trying to reproduce the problem on old versions of Solr/Lucene that aren't candidates for going into production. It's a pity. One thing that's bugging me is when I get errors in my log that look like this: <record> <date>2008-08-22T13:00:48</date> <millis>12194352486770</millis> <sequence>8387379</sequence> <logger>org.apache.solr.core.SolrCore</logger> <level>SEVERE</level> <class>org.apache.solr.common.SolrException</class> <method>log</method> <thread>56</thread> <message>java.lang.NullPointerException </message> </record> <record> <date>2008-08-22T13:00:48</date> <millis>1219435248676</millis> <sequence>8387381</sequence> <logger>org.apache.solr.servlet.SolrDispatchFilter</logger> <level>SEVERE</level> <class>org.apache.solr.common.SolrException</class> <method>log</method> <thread>56</thread> <message>java.lang.NullPointerException </message> </record> I get these NullPointException records every once and a while, always from SolrCore and SolrDispatchFilter. Don't get a stack trace, and no nearby errors seem to clarify what might have happened. On Thu, Aug 21, 2008 at 3:52 PM, Michael McCandless <[EMAIL PROTECTED]> wrote: > > OK indeed that revision of Lucene is before the workaround for that nasty > JRE bug was committed. > > Can you test one of those JRE versions (known not to have this particular > JRE bug) and see if you can get the original "massive deletion" problem to > happen. I guess first try it without the infoStream change, since it's > possible the infoStream change prevented the issue from happening? > > Mike > > Chris Harris wrote: > >> I'll see about using a newer/older JVM. >> >> In the meantime, according to the Solr admin page, which seems to get >> its info like so >> >> LucenePackage.class.getPackage().getImplementationVersion() >> >> what I've been testing is Lucene r652650. The Solr version is r654965, >> now modified of course to do some more debug logging. Unless I've >> screwed things up, this is the version of Lucene that comes with this >> version of Solr. >> >> On Thu, Aug 21, 2008 at 12:49 PM, Michael McCandless >> <[EMAIL PROTECTED]> wrote: >>> >>> Urgh, I was hoping we could repro the "massive deletion" with infoStream >>> turned on. >>> >>> Uh-oh: that "off by 1" corruption is very likely due to the Sun JRE bug >>> described here: >>> >>> https://issues.apache.org/jira/browse/LUCENE-1282 >>> >>> Can you downgrade to 1.6.0_03, or, upgrade to the latest beta build (b28) >>> of >>> Sun's 1.6 JRE, from here: >>> >>> http://download.java.net/jdk6/binaries/ >>> >>> I agree you should stick with the older stuff to reproduce this bug, but, >>> do >>> you know which svn rev of Lucene's JAR you are using? I had committed a >>> workaround for LUCENE-1282 to trunk. >>> >>> Mike > >
