[
https://issues.apache.org/jira/browse/HBASE-21213?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16631375#comment-16631375
]
stack commented on HBASE-21213:
-------------------------------
An opinion please [~allan163].
In my testing doing stuff like purging all MasterProcWALs testing hbck2 fixup,
I've manufactured a few odd cases where I want to be able to bypass a procedure
even though it has children: i.e. in PE, I'd add something like this:
if (!force) {
if (procedure.hasChildren()) {
LOG.info("{} has children, skipping bypass", procedure);
return false;
}
} else {
LOG.info("Bypassing child check!");
}
Cases are a MoveProcedure that has a lock on a region but its UnassignProcedure
is no longer in the record pushed out because millions of procedures have
passed through the system since. In the meantime, this stuck MoveProcedure is
making it so master proc wals are starting to backup.
I can figure why it got stuck and fix the issue later but it illustrates a case
where in hbck2 I will want to bypass a procedure even though it has children
unfinished supposedly. It is dangerous bypassing such a procedure since it
could make for hanging procedures... but in some cases I need to be able to do
it. I should probably add a special flag... Just wondering if you have any
thing you'd add here. Thanks.
> [hbck2] bypass leaves behind state in RegionStates when assign/unassign
> -----------------------------------------------------------------------
>
> Key: HBASE-21213
> URL: https://issues.apache.org/jira/browse/HBASE-21213
> Project: HBase
> Issue Type: Bug
> Components: amv2, hbck2
> Reporter: stack
> Assignee: stack
> Priority: Major
> Fix For: 2.1.1
>
> Attachments: HBASE-21213.branch-2.1.001.patch,
> HBASE-21213.branch-2.1.002.patch, HBASE-21213.branch-2.1.003.patch,
> HBASE-21213.branch-2.1.004.patch, HBASE-21213.branch-2.1.005.patch,
> HBASE-21213.branch-2.1.006.patch, HBASE-21213.branch-2.1.007.patch,
> HBASE-21213.branch-2.1.007.patch
>
>
> This is a follow-on from HBASE-21083 which added the 'bypass' functionality.
> On bypass, there is more state to be cleared if we are allow new Procedures
> to be scheduled.
> For example, here is a bypass:
> {code}
> 2018-09-20 05:45:43,722 INFO org.apache.hadoop.hbase.procedure2.Procedure:
> pid=100449, state=RUNNABLE:REGION_TRANSITION_DISPATCH, locked=true,
> bypass=LOG-REDACTED UnassignProcedure table=hbase:namespace,
> region=37cc206fe9c4bc1c0a46a34c5f523d16,
> server=ve1233.halxg.cloudera.com,22101,1537397961664 bypassed, returning null
> to finish it
> 2018-09-20 05:45:44,022 INFO
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor: Finished pid=100449,
> state=SUCCESS, bypass=LOG-REDACTED UnassignProcedure table=hbase:namespace,
> region=37cc206fe9c4bc1c0a46a34c5f523d16,
> server=ve1233.halxg.cloudera.com,22101,1537397961664 in 2mins, 7.618sec
> {code}
> ... but then when I try to assign the bypassed region later, I get this:
> {code}
> 2018-09-20 05:46:31,435 WARN
> org.apache.hadoop.hbase.master.assignment.RegionTransitionProcedure: There is
> already another procedure running on this region this=pid=100450,
> state=RUNNABLE:REGION_TRANSITION_QUEUE, locked=true; AssignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16
> owner=pid=100449, state=SUCCESS, bypass=LOG-REDACTED UnassignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16,
> server=ve1233.halxg.cloudera.com,22101,1537397961664 pid=100450,
> state=RUNNABLE:REGION_TRANSITION_QUEUE, locked=true; AssignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16; rit=OPENING,
> location=ve1233.halxg.cloudera.com,22101,1537397961664
> 2018-09-20 05:46:31,510 INFO
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor: Rolled back pid=100450,
> state=ROLLEDBACK,
> exception=org.apache.hadoop.hbase.procedure2.ProcedureAbortedException via
> AssignProcedure:org.apache.hadoop.hbase.procedure2.ProcedureAbortedException:
> There is already another procedure running on this region this=pid=100450,
> state=RUNNABLE:REGION_TRANSITION_QUEUE, locked=true; AssignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16
> owner=pid=100449, state=SUCCESS, bypass=LOG-REDACTED UnassignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16,
> server=ve1233.halxg.cloudera.com,22101,1537397961664; AssignProcedure
> table=hbase:namespace, region=37cc206fe9c4bc1c0a46a34c5f523d16
> exec-time=473msec
> {code}
> ... which is a long-winded way of saying the Unassign Procedure still exists
> still in RegionStateNodes.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
