[
https://issues.apache.org/jira/browse/HBASE-21844?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16761149#comment-16761149
]
Sergey Shelukhin commented on HBASE-21844:
------------------------------------------
As mentioned in HBASE-21797 I think it's quite reasonable to try to make AMv2
more resilient.
The common problem can be that proc WAL simply was not flushed due to an HDFS
issue, and master should be able to work around that if it actually has enough
state (which it does as illustrated here and HBASE-21742, SCP seems fully
redundant).
Esp. given how many issues we find... it's not realistic to run HBCK every day
to keep cluster alive which would be the case on any large cluster on 3.0 (or
2.2 which shares most of the assignment logic) right now in our experience.
> Master could get stuck in initializing state while waiting for meta
> -------------------------------------------------------------------
>
> Key: HBASE-21844
> URL: https://issues.apache.org/jira/browse/HBASE-21844
> Project: HBase
> Issue Type: Bug
> Components: master, meta
> Affects Versions: 3.0.0
> Reporter: Bahram Chehrazy
> Assignee: Bahram Chehrazy
> Priority: Major
> Attachments:
> 0001-HBASE-21844-Handling-incorrect-Meta-state-on-Zookeep.patch
>
>
> If the active master crashes after meta server dies, there is a slight chance
> of master getting into a state where the ZK says meta is OPEN, but the server
> is dead and there is no active SCP to recover it (perhaps the SCP has aborted
> and the procWALs were corrupted). In this case the waitForMetaOnline never
> returns.
>
> We've seen this happening a few times when there had been a temporary HDFS
> outage. Following log lines shows this state.
>
> 2019-01-17 18:55:48,497 WARN [master/************:16000:becomeActiveMaster]
> master.HMaster: hbase:meta,,1.1588230740 is NOT online; state=
> {1588230740 *state=*OPEN**, ts=1547780128227,
> server=*************,16020,1547776821322}
> ; *ServerCrashProcedures=false*. Master startup cannot progress, in
> holding-pattern until region onlined.
>
> I'm still investigating why and how to prevent getting into this bad state,
> but nevertheless the master should be able to recover during a restart by
> initiating a new SCP to fix the meta.
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
