jackrabbit-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Marcel Reutegger <marcel.reuteg...@gmx.net>
Subject Re: Index/data corruption
Date Wed, 25 Apr 2007 07:49:16 GMT
Hi Michael,

the stack trace indicate that jackrabbit is not properly shutdown. when 
jackrabbit is started up again the query handler runs recovery code as can be 
seen in the stack trace:

Michael Neale wrote:
>    at org.apache.jackrabbit.core.query.lucene.Recovery.run(Recovery.java
> :160)
>    at org.apache.jackrabbit.core.query.lucene.Recovery.run(Recovery.java
> :85)
>    at org.apache.jackrabbit.core.query.lucene.MultiIndex.<init>(
> MultiIndex.java:268)

the following IllegalArgumentException is probably a bug. I'll create a jira 
issue and fix it.

> java.lang.IllegalArgumentException: already contains: _c
>    at org.apache.jackrabbit.core.query.lucene.IndexInfos.addName(
> IndexInfos.java:170)
>    at org.apache.jackrabbit.core.query.lucene.MultiIndex.deleteIndex(
> MultiIndex.java:716)
>    at
> org.apache.jackrabbit.core.query.lucene.MultiIndex$DeleteIndex.execute(
> MultiIndex.java:1553)
>    at org.apache.jackrabbit.core.query.lucene.MultiIndex.executeAndLog(
> MultiIndex.java:809)
>    at org.apache.jackrabbit.core.query.lucene.MultiIndex.flush(
> MultiIndex.java:740)
>    at org.apache.jackrabbit.core.query.lucene.Recovery.run(Recovery.java
> :160)
[...]
> when trying to login to the repository the next time. This is extremely
> alarming to me. If I go in and blow away the contents of the index
> directories, it will then work for a while after that (once it has 
> recreted)
> but it only takes a dozen or so "cycles" (startup, run a query, 
> shutdown) to
> get back into this state.

Can you please make sure that Jackrabbit is shutdown properly in your cycles and 
check if you still see those exceptions? I could imagine that there's an issue 
in the index recovery code that only comes into play when one repeatedly kills a 
jackrabbit instance and then starts it again.

> This got me thinking, how are the lucene indexes kept in sync - they can't
> be part of the transaction can they? I mean they can try to be. This 
> worries
> me, I don't like being worried.

the lucene indexes are updated after a jackrabbit transaction is committed. if 
the jvm process is killed during that update, the index may be ouf of sync. See: 
http://issues.apache.org/jira/browse/JCR-204

this is a consequence of the decoupled design of the persistence manager and the 
query handler. one solution would be to have a two-phase commit for both changes 
to the persistence manager and the query handler. but this would require a more 
sophisticated interface for the persistence manager.

> Any advice much appreciated, this is pretty much stopping the show for me
> now.

The more details I see from your issue, the more I think it is caused by your 
test setup. Please make sure that you do a clean shutdown in your test cycle.

regards
  marcel

Mime
View raw message