Enis Soztutar created HBASE-8590:
------------------------------------
Summary: [0.94] BlockingMetaScannerVisitor should check for parent
meta entry while waiting for split daughter
Key: HBASE-8590
URL: https://issues.apache.org/jira/browse/HBASE-8590
Project: HBase
Issue Type: Bug
Components: regionserver
Reporter: Enis Soztutar
Assignee: Enis Soztutar
Fix For: 0.94.8
This was discovered after HBASE-8505 went in, which introduces a test
sporadically triggering this bug.
>From comments at HBASE-8505:
>From the logs at
>https://builds.apache.org/job/HBase-0.94-security/ws/trunk/target/surefire-reports/org.apache.hadoop.hbase.client.TestMetaScanner-output.txt,
> I think I understand what is going on:
BlockingMetaScannerVisitor blocks and wait for the split daughter to appear
when it sees a parent region (HBASE-5986). CatalogJanitor on the other hand
will order the regions in a (kind-of) topological sort (based on parent child
relation) so that it will guarantee parents are not GC'd before daughters.
What is happening in this issue is not related to the patch in this jira, but
the test triggers this extremely rare case by running concurrent
catalogjanitor, splits and metascanners. We have parent, splita and splitb
regions, and catalogjanitor decides to delete parent first and splitb in one
run. While there is a concurrent metascanner which will go over the parent, and
sees that it is split, but before being able to read the split daughter,
catalog janitor will delete both the parent and the child, which will lead to
metascanner blocking until timeout and failing the test.
On solution might be to also check whether the parent is still there in
BlockingMetaScannerVisitor while we are blocking for the daughter.
Good thing is that with HBASE-7721, we don't need any of this in trunk.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
