[
https://issues.apache.org/jira/browse/OAK-6915?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16244909#comment-16244909
]
Michael Dürig commented on OAK-6915:
------------------------------------
Wow, what an amazing find! Thanks [~dulceanu] and [~frm] for so much endurance.
AFAIU the core of the problem is {{SegmentCache.getSegment()}} always reading
from disk instead of trying the Guava cache first. A bit frightening that we
didn't notice this for so long in any of our tests.
The reason why reversing the statements in {{SegmentCache.put()}} (basically
undoing the
[changes|http://svn.apache.org/viewvc/jackrabbit/oak/trunk/oak-segment-tar/src/main/java/org/apache/jackrabbit/oak/segment/SegmentCache.java?r1=1793527&r2=1793526&pathrev=1793527]
from r1793527) seemingly fixes the problem is that in that case
{{SegmentId.unloaded()}} is not called for some segments that are evicted from
the Guava cache. Those segments will then stay memoized in the
{{SegmentId.segment}} field and thus will not need to be loaded from there on.
I say seemingly fixes the problem as reversing those statements back is not the
right solution IMO as in this case we introduce a memory leak in the
{{SegmentId}} class. This is what the comment in {{SegmentCache.put()}} hints
at: putting an item into the Guava cache can cause it to be evicted right away.
Even before the put method returns. If we call {{SegmentId.loaded()}} after the
call to {{SegmentCache.put()}}, we'll never ever get the corresponding call to
{{SegmentId.unloaded()}} any more, hence a leak.
That behaviour of the Guava cache can easily be reproduced in the debugger by
setting the cache size to 1M and observing the sequence of the calls to
{{SegmentCache.put()}} and {{SegmentCache.onRemove()}}. The striping of the
cache (16.stripes) actually exacerbates the issue as each stripe causes an
eviction once its size overflows 1/16th of the total size of the cache.
> Minimize the amount of uncached segment reads
> ---------------------------------------------
>
> Key: OAK-6915
> URL: https://issues.apache.org/jira/browse/OAK-6915
> Project: Jackrabbit Oak
> Issue Type: Improvement
> Components: segment-tar
> Reporter: Francesco Mari
> Assignee: Francesco Mari
> Fix For: 1.8, 1.7.12
>
> Attachments: OAK-6915-01.patch, OAK-6915-02.patch,
> OAK-6915-diagnostics.patch
>
>
> The current implementation of {{SegmentCache}} should make better use of the
> underlying Guava cache by relying on the cached segments instead of
> unconditionally performing an uncached segment read via the
> {{Callable<Segment>}} passed to {{SegmentCache#getSegment}}.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
