[
https://issues.apache.org/jira/browse/OAK-6500?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Chetan Mehrotra updated OAK-6500:
---------------------------------
Attachment: OAK-6500-ref-count-v1.patch
[patch|^OAK-6500-ref-count-v1.patch] using ref counting approach i.e. #B above
Here we refactor existing {{IndexNode}} class to 2
* IndexNode - Interface which captures the "api" of existing IndexNode. This
ensures that code which made use of IndexNode does not change. Note that all
client code use {{IndexTracker}} to obtain {{IndexNode}} instances
* IndexNodeManager - This class now has the logic present in earlier IndexNode
class and is used by {{IndexTracker}} to obtain {{IndexNode}} instances and
pass them to client code
h3. Ref Counting
This refactoring was required as we now need to maintain some state on per
acquire call. IndexNodeManager and NRTIndex now rely on ref counting support in
built in
[IndexReader|https://lucene.apache.org/core/4_7_1/core/org/apache/lucene/index/IndexReader.html#incRef()]
to ensure that reader instance gets closes when all references have been
released. The counting is managed in following way
# At time of creation of IndexReader the count = (0 -> 1)
#* Normal reader instances created once when IndexNodeManager is created have
the count 1
#* IndexReader for NRTIndex are also 1 when created in NRTIndex
# IndexSearcher creation increments the count to 2 - IndexNodeManager creates
new IndexSearcher and passes them either MultiReader or normal IndexReader. At
that time (IndexNodeManager#createReader) the count is cremented once
# Upon each acquire of IndexNode the count is further incremented (2->3)
# Then upon each release of IndexNode the count is decremented by 1 (3 -> 2).
This compliments #3
# Then for each refresh when a new IndexSearcher is created the count is again
decremented (2 -> 1). This complements #2
# Then upon close
#* Normal reader instance are closed in IndexNodeManager#close. The close call
decrements the count to 0 (1 -> 0)
#* IndexReader for NRTIndex - The decrement happens in NRTIndex#createReader
call where if a new reader is created then decRef is called for previous reader
This approach ensures that reader ref count should eventually reach 0 and thus
get closed.
h3. Asserting RefCount
In addtion NRTIndex exposes a flag to assert that reader ref counts are zero by
the time NRTIndex instance is closed. This assert can be enabled via system
property {{oak.lucene.assertAllResourcesClosed}}. The test setups enable this
flag to check that all resources are getting cleaned up
[~catholicon] - This patch changes a core component IndexNode. Due to rename
the diff is tricky. So would be helpful if you can review it once
[~teofili] - Would also be helpful if you can review the patch wrt Lucene api
usage
> NRTIndex leaks file handles due to unclosed IndexReader
> -------------------------------------------------------
>
> Key: OAK-6500
> URL: https://issues.apache.org/jira/browse/OAK-6500
> Project: Jackrabbit Oak
> Issue Type: Bug
> Components: lucene
> Affects Versions: 1.6.0
> Reporter: Chetan Mehrotra
> Assignee: Chetan Mehrotra
> Priority: Critical
> Fix For: 1.8, 1.7.6, 1.6.4
>
> Attachments: OAK-6500-approach-a-v1.patch,
> OAK-6500-ref-count-v1.patch, OAK-6500-v1.patch
>
>
> On some setups under stress it has been seen that NRTIndex leaks file handles
> over time.
> Checking with lsof indicates that more than 3 nrt folders per index are being
> used. However per design there can be max 3 and after system is not in use
> max 1 should be present.
> {noformat}
> $ lsof -p 9550 | grep '/nrt' | gawk 'match($0,
> /.*crx-quickstart\/repository\/index\/(.*?)\/\_.*$/, m) { print m[1]; }' |
> sort | uniq
> cqPageLucene-1501065263331/nrt1501065335930
> cqPageLucene-1501065263331/nrt1501065374667
> cqPageLucene-1501065263331/nrt1501065392492
> cqPageLucene-1501065263331/nrt1501065440955
> cqPageLucene-1501065263331/nrt1501065473286
> cqPageLucene-1501065263331/nrt1501065507345
> slingeventJob-1501065263330/nrt1501065356975
> slingeventJob-1501065263330/nrt1501065373229
> slingeventJob-1501065263330/nrt1501065394142
> slingeventJob-1501065263330/nrt1501065440953
> slingeventJob-1501065263330/nrt1501065473282
> slingeventJob-1501065263330/nrt1501065507342
> versionStoreIndex-1501065263332/nrt1501065335925
> versionStoreIndex-1501065263332/nrt1501065366781
> versionStoreIndex-1501065263332/nrt1501065392490
> versionStoreIndex-1501065263332/nrt1501065441232
> versionStoreIndex-1501065263332/nrt1501065473285
> {noformat}
> Further actually checking index folder indicates that those folder are
> actually deleted. So some where the file handle is still referring them.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
