[
https://issues.apache.org/jira/browse/HBASE-21073?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16647061#comment-16647061
]
Mike Drob commented on HBASE-21073:
-----------------------------------
bq. What does this do (smile)?
Right now, if there is a boolean configured in the hbase-site.xml then when
master comes up it will:
# host regions for system tables (on master)
# not assign any user regions
# skip setting up a few of the chores
All of this work is completely disjoint from the earlier maintenance mode,
which is why I gave it a different name. I didn't look too deeply into what old
maintenance mode did.
bq. Should maintenance mode bring up the Master but only have it listening for
Connection from localhost?
Maybe? Limiting the access is more work, and I'm not sure what it gains us.
bq. Is that it? This bit of code here is strange. In usual case – no regions on
master – we actually seem to hang here... which doesn't seem right.
It hangs... Yea, there's some race here when we need meta on master since we
won't mark the master as active until meta is hosted, but the region server
half of the same server won't come up until there's an active master. Normally
a region server would be able to pick up meta without waiting for an active
master to register with.
Going to hack around this by skipping the wait entirely. A little nervous about
it because I don't fully understand it, but I think it will be ok.
bq. its setting a running master into a 'state' so hbck1 could run.... and then
restoring old state after.
Gotcha. Will look at this for the next revision.
> "Maintenance mode" master
> -------------------------
>
> Key: HBASE-21073
> URL: https://issues.apache.org/jira/browse/HBASE-21073
> Project: HBase
> Issue Type: Sub-task
> Components: amv2, hbck2, master
> Reporter: stack
> Assignee: Mike Drob
> Priority: Major
> Attachments: HBASE-21073.master.001.patch,
> HBASE-21073.master.002.patch, HBASE-21073.master.003.patch
>
>
> Make it so we can bring up a Master in "maintenance mode". This is parse of
> master wal procs but not taking on regionservers. It would be in a state
> where "repair" Procedures could run; e.g. a Procedure that could recover meta
> by looking for meta WALs, splitting them, dropping recovered.edits, and even
> making it so meta is readable. See parent issue for why needed (disaster
> recovery).
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
