It's a usual practice in Solr to overwrite synonyms/stopwords files and issue a core reload command to force Solr reload these files. We've noticed that this trick is no longer working on trunk.
I started debugging this problem in Eclipse and I can see that Solr actually re-reads updated synonym file on core reload, but indexing process still uses old reference to Synonym filter / SynonymMap instance.
When I start Solr initially I can see that my Solr server has 2 instances of SynonymMap (the object that holds actual synonyms data). This is expected as I have 2 SynonymFilters defined in my schema (one field with 2 synonym filters - index + query time)
After a core reload I see that Solr has 4 instances of this class. So I thought that there could be some leak and issued many (N) core reload commands and expected to see 2*N instances. It didn't happen. I can only see 20 instances of this class.
This is pretty interesting number as well because, according to thread dump/list, Solr/Jetty has 10 working threads in thread pool (by default for example configs). So I suspect that we cache something in thread local storage and it hits us.
I looked into the code and found that Lucene caches token streams in ThreadLocal, but I don't know the code enough to state that this is the problem.
So I took a different approach and found the commit that introduced this bug. I wrote a simple test (shell script) and used git bisect tool to chase this bug.
ffd9c717448eca895d19be8ee9718bc399ac34a7 is the first bad commit
Author: Mark Robert Miller <firstname.lastname@example.org>
Date: Thu Jun 30 13:59:59 2011 +0000
SOLR-2193, SOLR-2565: The default Solr update handler has been improved so
that it uses fewer locks, keeps the IndexWriter open rather than closing it
on each commit (ie commits no longer wait for background merges to complete),
works with SolrCore to provide faster 'soft' commits, and has an improved API
that requires less instanceof special casing.
You may now specify a 'soft' commit when committing. This will
use Lucene's NRT feature to avoid guaranteeing documents are on stable storage in exchange
for faster reopen times. There is also a new 'soft' autocommit tracker that can be
SolrCores now properly share IndexWriters across SolrCore reloads.
git-svn-id: https://svn.apache.org/repos/asf/lucene/dev/trunk@1141542 13f79535-47bb-0310-9956-ffa450edef68
It looks like sharing IndexWriters across SolrCore reloads could be the root cause, right?