[
https://issues.apache.org/jira/browse/SOLR-1042?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ryuuichi Kumai updated SOLR-1042:
---------------------------------
Attachment: SOLR-1042.patch
> Memory leaks in DIH
> -------------------
>
> Key: SOLR-1042
> URL: https://issues.apache.org/jira/browse/SOLR-1042
> Project: Solr
> Issue Type: Bug
> Components: contrib - DataImportHandler
> Affects Versions: 1.3
> Reporter: Ryuuichi Kumai
> Attachments: SOLR-1042.patch
>
>
> If delta-import is executed many times, the heap utilization grows up and
> finally OutOfMemoryError occurs.
> When delta-import is executed with SqlEntityProcessor, the instances of
> TemplateString cached in VariableResolverImpl#TEMPLATE_STRING#cache.
> If the deltaQuery contains variable like `last_index_time', the cached values
> never used increases.
> Similarly, I guess that the cache increases when fetching each modified row
> with primary key.
> I think these queries should not be cached.
> I came up with two solutions:
> 1) Not to cache queries to get modified rows.
> 2) Make VariableResolverImpl#TEMPLATE_STRING non-static. Or clear cache on
> finishing delta-import.
> I think that #1 is better for performance than #2, but #2 is easier to solve
> the problem.
> I made a patch in #2 way, and then tested two solr applications with
> `-XX:+PrintClassHistgram' option.
> The result after importing several million rows from a MySQL database is as
> follows:
> * original solr-1.3:
> num #instances #bytes class name
> ----------------------------------------------
> ...
> 6: 2983024 119320960
> org.apache.solr.handler.dataimport.TemplateString
> ...
> * patched solr-1.3:
> num #instances #bytes class name
> ----------------------------------------------
> ...
> 748: 3 120
> org.apache.solr.handler.dataimport.TemplateString
> ...
> Though it is version 1.3 that I tested, perhaps current nightly version has
> same problem.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
