[
https://issues.apache.org/jira/browse/HBASE-20706?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16508310#comment-16508310
]
Josh Elser commented on HBASE-20706:
------------------------------------
bq. I'm saying that the change to MTP looks like a bug fix, not a hack.
Gotcha. I was being pedantic -- I wouldn't be surprised if some other changes
in-flight would invalidate the need for this bug-fix. That's what I meant by
calling this a hack.
bq. Does MTP get table lock for the life of the procedure? I thought it didn't.
You're correct, it does not grab a lock.
bq. I guess this is not safe for reopening a region. If a region is in OPENING
state, then it may have already loaded everything at the RS side, so if we do
not reopen it then it will miss the new configurations which have just
modified...
[~Apache9], thanks for looking.
Agreed that this fundamentally doesn't work for how we're using it and that we
want to do some kind of ReopenRegionProcedure for ModifyTableProcedure and
others.
I'm not sure if you agree with the fix I've made specifically for how MRP works
now to avoid this problem.
> [hack] Don't add known not-OPEN regions in reopen phase of MTP
> --------------------------------------------------------------
>
> Key: HBASE-20706
> URL: https://issues.apache.org/jira/browse/HBASE-20706
> Project: HBase
> Issue Type: Sub-task
> Components: amv2
> Reporter: Josh Elser
> Assignee: Josh Elser
> Priority: Critical
> Fix For: 3.0.0, 2.1.0, 2.0.1
>
> Attachments: HBASE-20706.001.branch-2.0.patch
>
>
> Shake-down of ModifyTableProcedure, talked this one out with Stack – "proper"
> fix is likely pending in HBASE-20682. Using MoveRegionProcedure is likely the
> wrong construct, we would want something specific to reopen (e.g. a
> ReopenProcedure).
> However, we're in a really bad state right now. If there are non-open regions
> for a table which has a modify submitted against it, the entire system locks
> up in a fast-spin while holding the table's lock. This fills up HDFS with PV2
> wals, and prevents you from doing anything in the hbase shell to try to fix
> those unassigned regions. You'll see spam in the master log like:
> {noformat}
> 2018-06-07 03:21:29,448 WARN [PEWorker-1] procedure.ModifyTableProcedure:
> Retriable error trying to modify table=METRIC_RECORD_HOURLY_UUID (in
> state=MODIFY_TABLE_REOPEN_ALL_REGIONS)
> org.apache.hadoop.hbase.client.DoNotRetryRegionException:
> a3dc333606d38aeb6e2ab4b94233cfbc is not OPEN
> at
> org.apache.hadoop.hbase.master.procedure.AbstractStateMachineTableProcedure.checkOnline(AbstractStateMachineTableProcedure.java:193)
> at
> org.apache.hadoop.hbase.master.assignment.MoveRegionProcedure.<init>(MoveRegionProcedure.java:67)
> at
> org.apache.hadoop.hbase.master.assignment.AssignmentManager.createMoveRegionProcedure(AssignmentManager.java:767)
> at
> org.apache.hadoop.hbase.master.assignment.AssignmentManager.createReopenProcedures(AssignmentManager.java:705)
> at
> org.apache.hadoop.hbase.master.procedure.ModifyTableProcedure.executeFromState(ModifyTableProcedure.java:128)
> at
> org.apache.hadoop.hbase.master.procedure.ModifyTableProcedure.executeFromState(ModifyTableProcedure.java:50)
> at
> org.apache.hadoop.hbase.procedure2.StateMachineProcedure.execute(StateMachineProcedure.java:184)
> at
> org.apache.hadoop.hbase.procedure2.Procedure.doExecute(Procedure.java:850)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.execProcedure(ProcedureExecutor.java:1472)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.executeProcedure(ProcedureExecutor.java:1240)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.access$800(ProcedureExecutor.java:75)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor$WorkerThread.run(ProcedureExecutor.java:1760)
> {noformat}
> We unstuck out internal test cluster giving the following change on top of
> Sergey's HBASE-20657. When choosing the regions to reopen, if we filter out a
> table's regions to only be those which are currently OPEN. There may be some
> transient failures here as well, but a subsequent retry of the reopen step
> should filter out that change.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
