[
https://issues.apache.org/jira/browse/HBASE-28420?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17823838#comment-17823838
]
Duo Zhang commented on HBASE-28420:
-----------------------------------
Confirmed, this is what we have in SplitWALRemoteProcedure.complete method
{code}
@Override
protected void complete(MasterProcedureEnv env, Throwable error) {
if (error == null) {
try {
env.getMasterServices().getSplitWALManager().archive(walPath);
} catch (IOException e) {
LOG.warn("Failed split of {}; ignore...", walPath, e);
}
succ = true;
} else {
if (error instanceof DoNotRetryIOException) {
LOG.warn("Sent {} to wrong server {}, try another", walPath,
targetServer, error);
succ = true;
} else {
LOG.warn("Failed split of {}, retry...", walPath, error);
succ = false;
}
}
}
{code}
We do not persist anything into the procedure store, so even the new active
master has already finished fencing, the old master could still accept the
report, and cause the new active master does not know the finishing of a
procedure.
This should be a common problem for remote procedures other than region
assigning, we should have a unified solution to fix them all.
Thanks.
> Aborting Active HMaster is not rejecting remote Procedure Reports
> -----------------------------------------------------------------
>
> Key: HBASE-28420
> URL: https://issues.apache.org/jira/browse/HBASE-28420
> Project: HBase
> Issue Type: Bug
> Components: master, proc-v2
> Affects Versions: 2.5.7
> Reporter: Umesh Kumar Kumawat
> Assignee: Umesh Kumar Kumawat
> Priority: Critical
>
> When the Active Hmaster is in the process of abortion and another HMaster is
> becoming Active HMaster,at the same time if any region server reports the
> completion of the remote procedure, it generally goes to the old active
> HMaster because of the cached value of rssStub ->
> [code|https://github.com/apache/hbase/blob/branch-2.5/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java#L2829]
> ([caller
> method|https://github.com/apache/hbase/blob/branch-2.5/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java#L3941]).
> On the Master side
> ([code|https://github.com/apache/hbase/blob/branch-2.5/hbase-server/src/main/java/org/apache/hadoop/hbase/master/MasterRpcServices.java#L2381]),
> It did check if the service is started but that returns true if the master
> is in the process of abortion(I didn't see when we are setting this flag
> false while abortion).
> This issue becomes *critical* when *ServerCrash of meta hosting RS and master
> failover* happens at the same time and hbase:meta got stuck in the offline
> state.
> Logs for abortion start of HMaster
> {noformat}
> 2024-02-02 07:33:11,581 ERROR [PEWorker-6] master.HMaster - ***** ABORTING
> master server4-1xxx,61000,1705169084562:
> FAILED persisting region=52d36581218e00a2668776cfea897132 state=CLOSING
> *****{noformat}
> {noformat}
> 2024-02-02 07:33:40,999 INFO [master/server4-1xxx:61000]
> regionserver.HRegionServer - Exiting;
> stopping=hbase2b-mnds4-1-ia2.ops.sfdc.net,61000,1705169084562; zookeeper
> connection closed.{noformat}
> it took almost 30 seconds to abort the HMaster.
>
> Logs of starting SCP for meta carrying host. (This SCP is started by the new
> active HMaster)
> {noformat}
> 2024-02-02 07:33:32,622 INFO [aster/server3-1xxx61000:becomeActiveMaster]
> assignment.AssignmentManager - Scheduled
> ServerCrashProcedure pid=3305546 for server5-1xxx61020,1706857451955
> (carryingMeta=true) server5-1-
> xxx61020,1706857451955/CRASHED/regionCount=1/lock=java.util.concurrent.locks.ReentrantReadWriteLock@1b0a5293[Write
>
> locks = 1, Read locks = 0], oldState=ONLINE.{noformat}
> initialization of remote procedure
> {noformat}
> 2024-02-02 07:33:33,178 INFO [PEWorker-4] procedure2.ProcedureExecutor -
> Initialized subprocedures=[{pid=3305548,
> ppid=3305547, state=RUNNABLE; SplitWALRemoteProcedure server5-1-
> xxxxt%2C61020%2C1706857451955.meta.1706858156058.meta,
> worker=server4-1-xxxx,61020,1705169180881}]{noformat}
> Logs of remote procedure handling on Old Active Hmaster(server4-1xxx,61000)
> (in the process of abortion)
> {noformat}
> 2024-02-02 07:33:37,990 DEBUG
> [r.default.FPBQ.Fifo.handler=243,queue=9,port=61000] master.HMaster - Remote
> procedure
> done, pid=3305548{noformat}
> This should be handled by the new active HMaster so that it can wake up the
> suspended Procedure on the new Active Hmaster. As the new ActiveHMaster was
> not able to wake that up, SCP procedure got stuck thus meta stayed OFFLINE.
>
> Logs of Hmaster trying to becomeActivehmaster but stuck-
> {noformat}
> 2024-02-02 07:33:43,159 WARN [aster/server3-1-ia2:61000:becomeActiveMaster]
> master.HMaster - hbase:meta,,1.1588230740
> is NOT online; state={1588230740 state=OPEN, ts=1706859212481,
> server=server5-1-xxx,61020,1706857451955};
> ServerCrashProcedures=true. Master startup cannot progress, in
> holding-pattern until region onlined.{noformat}
> After this master was stuck till we did hmaster failover to come out of this
> situation.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
