[ https://issues.apache.org/jira/browse/OAK-4732?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16074435#comment-16074435 ]
Michael Dürig commented on OAK-4732: ------------------------------------ Patch looks good to me. Maybe we increase the {{SynchronizedDescriptiveStatistics}} window a bit? 1000 instead of 100? As a follow up, it would be good to add a feature flag to control this feature if this turns out to be necessary. One way would be to add a flag to configure the percentile used in {{LockBasedScheduler.getHeadNodeState()}}. Setting it to {{0}} would basically disable this. Setting it to a different value might be interesting in future experiments. > (Slightly) prioritise reads over writes > ---------------------------------------- > > Key: OAK-4732 > URL: https://issues.apache.org/jira/browse/OAK-4732 > Project: Jackrabbit Oak > Issue Type: Improvement > Components: segment-tar > Reporter: Michael Dürig > Assignee: Andrei Dulceanu > Priority: Minor > Labels: Performance, scalability > Fix For: 1.8, 1.7.4 > > Attachments: OAK-4732.patch > > > When fetching the current root from the {{SegmentNodeStore}} an older > revision will be returned when a commit is being processed concurrently. I > think it would make sense to wait for a short time in this case increasing > the chance of returning an up to date state. The idea is that this would > lower the rebasing work that need to be done later on should the returned > root be used for further modifications. > An interesting value for the wait time is to use the median (or more general > a percentile) of the commit time of the last say 1000 commits. This would > mean that (for the median) we have a 50% chance of getting up to date date. > For a 90% percentile we would have longer wait times but then a 90% chance of > getting up to date date. -- This message was sent by Atlassian JIRA (v6.4.14#64029)