[
https://issues.apache.org/jira/browse/HBASE-20792?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16524067#comment-16524067
]
Josh Elser commented on HBASE-20792:
------------------------------------
The regions that SCP will try to do splitting for are defined by what it sees
in SERVER_CRASH_GET_REGIONS. This is ultimately pulled from hbase:meta that is
populated when AM starts up. In other words, AM didn't tell SCP that these
Regions needed to be recovered (SCP just assumes that they must be out). This
lines up with what meta looks like (going off of the meta output in the
description).
Specifically, in the above example where we can info:server=rs12, info:sn=rs6,
and info:state=OPEN for that region, what we see for the hbase:namespace would
be:
* info:server=rs5
* info:sn=rs6
* info:state=OPEN
If we had an SCP submitted for RS5, that would've "fixed" this for us, but that
successfully completed *before* the master was restarted.
I am near positive the question to answer is: how did info:server get set back
to info:server=rs5?
> info:servername and info:sn inconsistent for OPEN region
> --------------------------------------------------------
>
> Key: HBASE-20792
> URL: https://issues.apache.org/jira/browse/HBASE-20792
> Project: HBase
> Issue Type: Bug
> Components: Region Assignment
> Reporter: Josh Elser
> Assignee: Josh Elser
> Priority: Blocker
> Fix For: 2.0.2
>
>
> Next problem we've run into after HBASE-20752 and HBASE-20708
> After a rolling restart of a cluster, we'll see situations where a collection
> of regions will simply not be assigned out to the RS. I was able to reproduce
> this my mimic the restart patterns our tests do internally (ignore whether
> this is the best way to restart nodes for now :)). The general pattern is
> this:
> {code:java}
> for rs in regionservers:
> stop(server, rs, RS)
> for master in masters:
> stop(server, master, MASTER)
> sleep(15)
> for master in masters:
> start(server, master, MASTER)
> for rs in regionservers:
> start(server, rs, RS){code}
> Looking at meta, we can see why the Master is ignoring some regions:
> {noformat}
> test
> column=table:state, timestamp=1529871718998, value=\x08\x00
> test,,1529871718122.0297f680df6dc0166a44f9536346268e.
> column=info:regioninfo, timestamp=1529967103390, value={ENCODED =>
> 0297f680df6dc0166a44f9536346268e, NAME =>
> 'test,,1529871718122.0297f680df6dc0166a44f9536346268e.', STARTKEY
> => '', ENDKEY =>
> ''}
> test,,1529871718122.0297f680df6dc0166a44f9536346268e.
> column=info:seqnumDuringOpen, timestamp=1529967103390,
> value=\x00\x00\x00\x00\x00\x00\x00*
> test,,1529871718122.0297f680df6dc0166a44f9536346268e.
> column=info:server, timestamp=1529967103390,
> value=ctr-e138-1518143905142-378097-02-000012.hwx.site:16020
> test,,1529871718122.0297f680df6dc0166a44f9536346268e.
> column=info:serverstartcode, timestamp=1529967103390, value=1529966776248
> test,,1529871718122.0297f680df6dc0166a44f9536346268e. column=info:sn,
> timestamp=1529967096482,
> value=ctr-e138-1518143905142-378097-02-000006.hwx.site,16020,1529966755170
> test,,1529871718122.0297f680df6dc0166a44f9536346268e.
> column=info:state, timestamp=1529967103390, value=OPEN{noformat}
> The region is marked as {{OPEN}}. The master doesn't know any better.
> However, the interesting bit is that {{info:server}} and {{info:sn}} are
> inconsistent (which, according to the javadoc should not be possible for an
> {{OPEN}} region).{{}}
> This doesn't happen every time, but I caught it yesterday on the 2nd or 3rd
> attempt, so I'm hopeful it's not a bear to repro.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
