[
https://issues.apache.org/jira/browse/HBASE-21259?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16635924#comment-16635924
]
stack edited comment on HBASE-21259 at 10/3/18 12:00 AM:
---------------------------------------------------------
[~allan163]
* meta has a region in CLOSING state against a server that has no mention in
fs, is not online, nor has it a znode so it is 'unknown' to the system.
* I try to move the region 'manually' via hbck2 from CLOSING to CLOSED -- i.e.
unassign -- so I can assign it elsewhere. The CLOSING dispatch fails because no
such server and the UP expires the server which queues a SCP.
* The SCP runs. Finds no logs to splits. Finds the stuck UP and calls it's
handleFailure. The UP then moves to CLOSED and all is good.
* Except, the SCP has now resulted in their being a deadserver element.
* So, when the next region that references the 'unknown' server comes along,
it goes to unassign, fails, and tries to queue a server expiration.....
* But the attempt at expiration is rejected because 'there is one in progress
already' (because the server has an entry in dead servers -- See
ServerManager#expireServer) so we skip out without queuing a new SCP.
* This second UP and all subsequent regions that were pointing to the
'unknown' server end up 'hung', suspended waiting for someone to wake them.
I have to call 'bypass' on each to get them out of suspend. I cannot unassign
the regions, not in bulk at least.
If a server is dead we should not be reviving it. It causes more headache that
it solves.
My first patch was stopping our reviving a server if it unknown but it messed
up startup. Let me try and be more clinical.
was (Author: stack):
[~allan163]
* meta has a region in CLOSING state against a server that has no mention in
fs, is not online, nor has it a znode so it is 'unknown' to the system.
* I try to move the region from CLOSING to CLOSED so I can assign it
elsewhere. The CLOSING dispatch fails because no such server and UP queues a
SCP.
* The SCP runs. Finds no logs to splits. Finds the stuck UP and calls its
handleFailure. The UP then moves to CLOSED and all is good.
* Except, the SCP has now resulted in their being a deadserver element.
* So, when the new region that references the 'unknown' server comes along, it
goes to unassign, fails, and tries to queue an expiration.
* But the attempt at expiration is rejected because 'there is one in progress
already' (because the server has an entry in dead servers) so we skip out
without queuing a new SCP.
* The second UP and all subsequent regions that were pointing to the 'unknown'
server end up 'hung', suspended waiting for someone to wake them.
If a server is dead we should not be reviving it. It cause more headache that
it solves.
My first patch was stopping our reviving a server if it unknown but it messed
up startup. Let me try and be more clinical.
> [amv2] Revived deadservers; recreated serverstatenode
> -----------------------------------------------------
>
> Key: HBASE-21259
> URL: https://issues.apache.org/jira/browse/HBASE-21259
> Project: HBase
> Issue Type: Bug
> Components: amv2
> Affects Versions: 2.1.0
> Reporter: stack
> Assignee: stack
> Priority: Major
> Fix For: 2.2.0, 2.1.1, 2.0.3
>
> Attachments: HBASE-21259.branch-2.1.001.patch
>
>
> On startup, I see servers being revived; i.e. their serverstatenode is
> getting marked online even though its just been processed by
> ServerCrashProcedure. It looks like this (in a patched server that reports on
> whenever a serverstatenode is created):
> {code}
> 2018-09-29 03:45:40,963 INFO
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor: Finished pid=3982597,
> state=SUCCESS; ServerCrashProcedure
> server=vb1442.halxg.cloudera.com,22101,1536675314426, splitWal=true,
> meta=false in 1.0130sec
> ...
> 2018-09-29 03:45:43,733 INFO
> org.apache.hadoop.hbase.master.assignment.RegionStates: CREATING!
> vb1442.halxg.cloudera.com,22101,1536675314426
> java.lang.RuntimeException: WHERE AM I?
> at
> org.apache.hadoop.hbase.master.assignment.RegionStates.getOrCreateServer(RegionStates.java:1116)
> at
> org.apache.hadoop.hbase.master.assignment.RegionStates.addRegionToServer(RegionStates.java:1143)
> at
> org.apache.hadoop.hbase.master.assignment.AssignmentManager.markRegionAsClosing(AssignmentManager.java:1464)
> at
> org.apache.hadoop.hbase.master.assignment.UnassignProcedure.updateTransition(UnassignProcedure.java:200)
> at
> org.apache.hadoop.hbase.master.assignment.RegionTransitionProcedure.execute(RegionTransitionProcedure.java:369)
> at
> org.apache.hadoop.hbase.master.assignment.RegionTransitionProcedure.execute(RegionTransitionProcedure.java:97)
> at
> org.apache.hadoop.hbase.procedure2.Procedure.doExecute(Procedure.java:953)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.execProcedure(ProcedureExecutor.java:1716)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.executeProcedure(ProcedureExecutor.java:1494)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor.access$900(ProcedureExecutor.java:75)
> at
> org.apache.hadoop.hbase.procedure2.ProcedureExecutor$WorkerThread.run(ProcedureExecutor.java:2022)
> {code}
> See how we've just finished a SCP which will have removed the
> serverstatenode... but then we come across an unassign that references the
> server that was just processed. The unassign will attempt to update the
> serverstatenode and therein we create one if one not present. We shouldn't be
> creating one.
> I think I see this a lot because I am scheduling unassigns with hbck2. The
> servers crash and then come up with SCPs doing cleanup of old server and
> unassign procedures in the procedure executor queue to be processed still....
> but could happen at any time on cluster should an unassign happen get
> scheduled near an SCP.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
