[
https://issues.apache.org/jira/browse/OAK-5602?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16096000#comment-16096000
]
Thomas Mueller commented on OAK-5602:
-------------------------------------
> periodically having each instance inform the rest
If an instance is a cluster node, what if this instance is stopped for a longer
time (a day or two). Maybe it will be started again, maybe not. I see two
options, both with a risk:
* A: If we assume this instance still needs to read the info, we possibly wait
forever, so never clean up.
* B: If we assume this instance doesn't need to read it because it didn't say
so for a long time, then we possibly remove data that is still needed.
Maybe we should use the "pessimistic" A route, with warnings in the log file
saying it didn't respond / update the lease, and description on how to
_manually_ remove the instance from the list.
> transaction is very long running
Maybe best would be to kill (rollback) such transactions after some time, that
is (x and y to be defined, and y > x):
* the session / instance that writes to the transaction stops writing and rolls
back after x minutes, and
* all other other sessions / instances assume it is stopped after y minutes
Otherwise I wouldn't know how to distinguish between case (a) the process was
killed, and (b) the process is still running but just slow.
> avoid missing journal entries
> -----------------------------
>
> Key: OAK-5602
> URL: https://issues.apache.org/jira/browse/OAK-5602
> Project: Jackrabbit Oak
> Issue Type: Improvement
> Components: core
> Affects Versions: 1.6.0
> Reporter: Stefan Egli
> Fix For: 1.8
>
>
> As a follow up of OAK-5601: that one is about a situation where
> backgroundRead falls way behind and journal GC cleans up journal entries
> before it was able to read it. We should generally look into improving this
> situation, eg by having journal GC not do the clean up as long as there are
> instances that haven't read the entry. This could perhaps be achieved by
> periodically having each instance inform the rest about what journal entries
> it has processed (or simpler: how far back it would be -find- fine for
> journal entries to be GC'ed)
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
