[
https://issues.apache.org/jira/browse/OAK-2673?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14390672#comment-14390672
]
Vikas Saurabh commented on OAK-2673:
------------------------------------
I tried a local run with/without fix where a script uploaded around 4000 assets
using 100 concurrent threads. I did that on a single instance cluster with
mongo back-end. While I couldn't reproduce a merge failure even without fix.
But trace level logging on DocumentNodeStoreBranch showed around 2500 conflicts
which got retried (I don't know if all 2500 got through in 1 retry or if there
were some commits which required multiple retries). Otoh, with fix that number
dropped significantly (~200) -- On broad level I think even those should have
got resolved. I'd try to look up if I missed something silly in current patch
or if there was an altogether different case that I missed.
> Resolve add-add, delete-delete merge conflict for empty hidden docs
> -------------------------------------------------------------------
>
> Key: OAK-2673
> URL: https://issues.apache.org/jira/browse/OAK-2673
> Project: Jackrabbit Oak
> Issue Type: Sub-task
> Components: mongomk
> Reporter: Vikas Saurabh
> Assignee: Marcel Reutegger
> Fix For: 1.2
>
> Attachments:
> OAK-2673-Adding-resolution-of-same-name-conflicts-ad(take2).patch,
> OAK-2673-HiddenNodeSameNameMergeConflict.patch
>
>
> While OAK-1550 is about general fix for resolving same node merge conflicts.
> But until general issue is fixed, we should special case for hidden nodes
> (e.g. index sub-tree).
> Discussed offline with [~mreutegg] and [~chetanm] about handling merge
> conflict for hidden nodes. Main concern to not resolve the conflict seemed to
> be observation (specifying/declaring what events reach which clients). For
> hidden nodes, observation isn't a concern. Along with that, in heavy write
> scenarios there are some cases (I'm aware of property index updates) which
> often cause conflicts -- thereby wasting some time during merge attempts.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
