[
https://issues.apache.org/jira/browse/HBASE-12856?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14277713#comment-14277713
]
Andrew Purtell commented on HBASE-12856:
----------------------------------------
Lars and I discussed this issue some more. Having a weak reference to the
classloader in the map is fine, the values will point to CoprocessorClassLoader
instances that should never be GCed because we don't unload coprocessor
classes. The CoprocessorClassLoader instances will be strongly referenced by
every java.lang.Class they instantiated, indefinitely.
Even then, we really don't want the "classLoadersCache" map in
CoprocessorClassLoader to be a cache, we want it to be an index. If/when we
implement coprocessor class unloading we wouldn't want a strong reference in
this map retaining the classloader. The comments around this field and even its
name could be reconsidered, 'cache' -> 'index'.
It seems fine to leave this as is. However I will leave this issue open for a
bit in case we are mistaken in some way or there is additional comment or
suggestion.
> Revisit table coprocessor classloader caching
> ---------------------------------------------
>
> Key: HBASE-12856
> URL: https://issues.apache.org/jira/browse/HBASE-12856
> Project: HBase
> Issue Type: Improvement
> Reporter: Andrew Purtell
> Assignee: Andrew Purtell
> Fix For: 2.0.0, 0.98.10, 1.1.0
>
>
> For table coprocessors, we create coprocessor classloaders and cache them
> keyed on the path to the coprocessor jar. However, we cache weak refs so it's
> quite possible if the refs are not being retained due to heap pressure we
> might create a new classloader on the next region open, i.e. after a split or
> relocation. If under very heavy write load, perhaps ingest with all edits
> skipping the WAL for example, we'd be both generating a ton of garbage and
> rapidly splitting at the same time.
> We should revisit the notion of keeping only weak refs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
