[
https://issues.apache.org/jira/browse/OAK-7859?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16683248#comment-16683248
]
Amit Jain commented on OAK-7859:
--------------------------------
[~wim.symons] Yes the {{blobTrackSnapshotIntervalInSecs}} controls the when the
snapshots are taken is partly responsible for the above i.e running DSGC again
after the snapshot has been taken might reduce this.
But there are other cases where the deleted references can get resurrected from
other instances [1], this was fixed in 1.6.11 and an update to that will help.
Also, in your case I think it will also help to manually delete all the .refs
files available in the bucket, then use collectGarbage(false, true) once and
then revert to the collectGarbage(false, false) once you are on the latest Oak
version.
[1] https://issues.apache.org/jira/browse/OAK-7209
> S3 Bucket iterator stops too early
> ----------------------------------
>
> Key: OAK-7859
> URL: https://issues.apache.org/jira/browse/OAK-7859
> Project: Jackrabbit Oak
> Issue Type: Bug
> Components: blob-cloud
> Affects Versions: 1.6.6
> Reporter: Wim Symons
> Assignee: Amit Jain
> Priority: Critical
> Labels: newbie, pull-request-available
> Attachments: META-ids.txt, Screenshot 2018-11-06 at 13.44.15.png,
> dsgc-bis.zip, dsgc.log, meta-info.txt
>
>
> Fixed a major bug in the S3 bucket iterator.
> When the returned queue of records is empty due to the fact that we get a
> full page of records starting with the META/ key, the iterator stops while
> there is still data available in the bucket.
> This causes problems with datastore GC, and datastore consistency checks
> (both online and offline), and possibly even more.
> A little explainer. But based on a batch size of 2 instead of 1000.
> Suppose your list of S3 keys looks as follows:
> * 1
> * 2
> * 3
> * 4
> * META/1
> * META/2
> * 5
> * 6
> loadBatch would first load [1, 2], filter out no META/ keys and pass [1, 2]
> to the caller.
> Next time, loadBatch would load [3, 4], filter out no META/ keys and pass [3,
> 4] to the caller.
> Than, loadBatch would load [META/1, META/2], filter out the META/ keys and
> pass [] to the caller.
> When that happens, traversing the bucket would stop, because the returned
> list is empty, even if there are many more batches to load.
> The fix checks if the returned list is empty and there are more batches
> available, it would load (a) new batch(es) until there is data in the batch
> or there is no more batch available.
> We are currently running Oak 1.6.6 on AEM 6.3.1.2, but as the bug is still in
> trunk, all previous versions of Oak are affected as well.
> I provided 2 pull requests: one for trunk
> ([https://github.com/apache/jackrabbit-oak/pull/103)] and one for the 1.6
> branch ([https://github.com/apache/jackrabbit-oak/pull/104).]
> CI failed on [https://github.com/apache/jackrabbit-oak/pull/103,] but I don't
> think it's related to my changes.
> For the record, the patch works as I was able to successfully test this on
> our production repository using oak-run --id. With version 1.6.6 it reported
> 800k items, with my patched version, it reported 1.8m items. (As our META/
> nodes are listed somewhere half-way through.)
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
