[
https://issues.apache.org/jira/browse/HBASE-18099?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16081448#comment-16081448
]
Gary Helmling commented on HBASE-18099:
---------------------------------------
bq. Is there suggestion on how backward compatibility can be kept ?
Sure, don't add the method to Region, just to HRegion, check for an instance of
HRegion in FlushSnapshotSubprocedure and cast the instance before calling the
method.
This is a critical issue that means that our current snapshot implementation in
1.3 is broken. It doesn't guarantee that all of the data that should be
present in a snapshot is there. I think all of the users of the impacted
versions would be interested in a fix.
> FlushSnapshotSubprocedure should wait for concurrent Region#flush() to finish
> -----------------------------------------------------------------------------
>
> Key: HBASE-18099
> URL: https://issues.apache.org/jira/browse/HBASE-18099
> Project: HBase
> Issue Type: Bug
> Reporter: Ted Yu
> Assignee: Ted Yu
> Priority: Critical
> Fix For: 2.0.0, 1.4.0
>
> Attachments: 18099.v1.txt, 18099.v2.txt, 18099.v3.txt, 18099.v4.txt
>
>
> In the following thread:
> http://search-hadoop.com/m/HBase/YGbbMXkeHlI9zo
> Jacob described the scenario where data from certain region were missing in
> the snapshot.
> Here was related region server log:
> https://pastebin.com/1ECXjhRp
> He pointed out that concurrent flush from MemStoreFlusher.1 thread was not
> initiated from the thread pool for snapshot.
> In RegionSnapshotTask#call() method there is this:
> {code}
> region.flush(true);
> {code}
> The return value is not checked.
> In HRegion#flushcache(), Result.CANNOT_FLUSH may be returned due to:
> {code}
> String msg = "Not flushing since "
> + (writestate.flushing ? "already flushing"
> : "writes not enabled");
> {code}
> This implies that FlushSnapshotSubprocedure may incorrectly skip waiting for
> the concurrent flush to complete.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
