Mailing-List: contact solr-user-help@lucene.apache.org; run by ezmlm
Precedence: bulk
Reply-To: solr-user@lucene.apache.org
Received-SPF: pass (athena.apache.org: domain of seb2lille@gmail.com
 designates 209.85.217.176 as permitted sender)
MIME-Version: 1.0
Sender: seb2lille@gmail.com
In-Reply-To: 
 <CAKzVNR5n2yrmybnRL_59VRNGNGdLN_b_KJ=s5NRzYxDdgP9WAA@mail.gmail.com>
References: 
 <CAKzVNR5k2pktX0QfgDa7R8vKvwyr0yWacaA9uW7ZX1WPq7uccQ@mail.gmail.com>
	<CALON7Or1k1XBvO3SR6DDkqmZdPKtFMTF3SEV7x_Eoqov2Zk4Xw@mail.gmail.com>
	<CAKzVNR5n2yrmybnRL_59VRNGNGdLN_b_KJ=s5NRzYxDdgP9WAA@mail.gmail.com>
Date: Thu, 9 Jan 2014 10:34:19 +0100
Message-ID: 
 <CALON7OpWS+5uvxSMcti+7Ck8nhMPZKqDSs-A1i95mPq67BOMeg@mail.gmail.com>
Subject: Re: solr OOM Crash
From: =?ISO-8859-1?Q?S=E9bastien_Michel?= <sebastien.michel@atos.net>
To: solr-user@lucene.apache.org
Content-Type: multipart/alternative; boundary=047d7b3434bc546f5f04ef8650f8

--047d7b3434bc546f5f04ef8650f8
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable

Hi Sandra,

Excuse me for the late reply.
We use lotsofcores (http://wiki.apache.org/solr/LotsOfCores) Solr feature,
around 100 simultaneous loaded cores. But the issue is reproducible with
few less cores.
We also have a high rate of indexing, and also reindexing (atomic update).

We are indexing media files metadata, but also metadata and contents of
PDF, the content is stored in a "text" field (stored=3D"true").
Until release 4.3, Solr  uses a growing buffer to uncompress stored fields
(I assume one buffer per Solr Core or per Shard).
The issue comes when Solr read some big docs, the buffer of
CompressedStoredFieldReader grows but never shrinks.  The more such big
docs are read in different threads, the more the Heap usage is growing,
until the heap has no more free memory available and GC runs continuously.

Analyzing the dump: Class byte[] takes 3gb out of 4gb resourced to the JVM,
mainly referenced by CompressingStoredFieldsReader

I hope it can help you.

S=E9bastien


2013/12/30 Sandra Scott <scottsandra85@gmail.com>

> Hello S=E9bastien,
>
> Can you give some information about your environment so I can make sure w=
e
> are having the same problem you had?
> Also, did you find out what caused the GC to go crazy or what caused the
> increased commit rate?
>
> Thanks,
> Sandra
>
>
> On Thu, Dec 19, 2013 at 12:34 PM, S=E9bastien Michel <
> sebastien.michel@atos.net> wrote:
>
> > Hi Sandra,
> >
> > I'm not sure if your problem is same as ours, but we encountered the sa=
me
> > issue on our Solr 4.2, the major memory usage was due to
> > CompressingStoredFieldsReader and GC became crazy.
> > In our context, we have some stored fields and for some documents the
> > content of the text field could be huge.
> >
> > We resolved our issue with the backport of this fix :
> > https://issues.apache.org/jira/browse/LUCENE-4995
> >
> > You should also upgrade to Solr 4.4 or more
> >
> > Regards,
> > S=E9bastien
> >
> >
> > 2013/12/12 Sandra Scott <scottsandra85@gmail.com>
> >
> > > Helllo,
> > >
> > > We are experiencing unexplained OOM crashes. We have already seen it =
a
> > few
> > > times, over our different solr instances. The crash happens only at a
> > > single shard of the collection.
> > >
> > > Environment details:
> > > 1. Solr 4.3, running on tomcat.
> > > 2. 24 Shards.
> > > 3. Indexing rate of ~800 docs per minute.
> > >
> > > Solrconfig.xml:
> > > 1. Merge factor 4
> > > 2. Sofrcommit every 10 min
> > > 3. Hardcommit every 30 min
> > >
> > > Main findings:
> > > 1. Solr logs: No query failures prior to the OOM, but DOUBLE the amou=
nt
> > of
> > > soft and hard commits in comparison to other shards.
> > > 2. Analyzing the dump (VisualVM): Class byte[] takes 4gb out of 5gb
> > > resourced to the JVM, mainly referenced by
> CompressingStoredFieldsReader
> > GC
> > > root (which by looking at the code, we suspect they were created due =
to
> > > CompressingSortedFieldsWriter.merge).
> > >
> > > Sub findings:
> > > 1. GC logs: Showed 108 GC fails prior to the crash.
> > > 2. CPI: Overall usage seems fine, but the % of CPU time for the GC
> stays
> > > high 6 min before the OOM.
> > > 3. Memory: Half an hour before OOM the usage slowly rises, until it
> gets
> > to
> > > 5.4gb.
> > >
> > > Has anyone encountered higher than normal commit rate that seem to
> > increase
> > > merge rate and cause what I described?
> > >
> >
>

--047d7b3434bc546f5f04ef8650f8--