[
https://issues.apache.org/jira/browse/OAK-1471?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13922979#comment-13922979
]
Tobias Bocanegra commented on OAK-1471:
---------------------------------------
result:
* concurrent writes are almost twice as fast for the lower concurrency range
(2-8) but also overall better. at least they show no degradation in through put
even for high concurrency.
* concurrent normal ACL reads are almost the same
* concurrent everyone reads are about 30% slower, for a test set of 10k
everyone ACEs
> Mod count in permission store requires cluster wide synchronization
> -------------------------------------------------------------------
>
> Key: OAK-1471
> URL: https://issues.apache.org/jira/browse/OAK-1471
> Project: Jackrabbit Oak
> Issue Type: Bug
> Components: core
> Reporter: Marcel Reutegger
> Assignee: Tobias Bocanegra
> Priority: Critical
> Fix For: 0.19
>
> Attachments: OAK-1471.patch
>
>
> The current permission store implementation keeps a mod count property on
> some of the nodes to detect changes and update/invalidate a cache when
> required. In general this introduces a contention point when there are
> concurrent modifications of access control entries for a given user. Those
> concurrent changes may introduce conflicts and commits may have to be retried
> after a rebase. This still works somewhat OK when there is just a single oak
> instance, because both segment and document store implementation will
> eventually fall back to locking and serialize the commits. However, it does
> not work well in a cluster, unless we apply the same locking cluster wide.
> See also discussion here:
> http://markmail.org/message/k64udmgc3ctaqmn2
--
This message was sent by Atlassian JIRA
(v6.2#6252)
