[jira] [Commented] (FLINK-34672) HA deadlock between JobMasterServiceLeadershipRunner and DefaultLeaderElectionService
[
https://issues.apache.org/jira/browse/FLINK-34672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17849916#comment-17849916
]
Hong Liang Teoh commented on FLINK-34672:
-
Ok - sounds good. Will proceed with 1.19.1 RC without this
> HA deadlock between JobMasterServiceLeadershipRunner and
> DefaultLeaderElectionService
> -
>
> Key: FLINK-34672
> URL: https://issues.apache.org/jira/browse/FLINK-34672
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
>Affects Versions: 1.17.2, 1.19.0, 1.18.1, 1.20.0
>Reporter: Chesnay Schepler
>Assignee: Matthias Pohl
>Priority: Major
> Labels: pull-request-available
> Fix For: 1.18.2, 1.20.0, 1.19.1
>
>
> We recently observed a deadlock in the JM within the HA system.
> (see below for the thread dump)
> [~mapohl] and I looked a bit into it and there appears to be a race condition
> when leadership is revoked while a JobMaster is being started.
> It appears to be caused by
> {{JobMasterServiceLeadershipRunner#createNewJobMasterServiceProcess}}
> forwarding futures while holding a lock; depending on whether the forwarded
> future is already complete the next stage may or may not run while holding
> that same lock.
> We haven't determined yet whether we should be holding that lock or not.
> {code}
> "DefaultLeaderElectionService-leadershipOperationExecutor-thread-1" #131
> daemon prio=5 os_prio=0 cpu=157.44ms elapsed=78749.65s tid=0x7f531f43d000
> nid=0x19d waiting for monitor entry [0x7f53084fd000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.runIfStateRunning(JobMasterServiceLeadershipRunner.java:462)
> - waiting to lock <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.revokeLeadership(JobMasterServiceLeadershipRunner.java:397)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.notifyLeaderContenderOfLeadershipLoss(DefaultLeaderElectionService.java:484)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1252/0x000840ddec40.accept(Unknown
> Source)
> at java.util.HashMap.forEach(java.base@11.0.22/HashMap.java:1337)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.onRevokeLeadershipInternal(DefaultLeaderElectionService.java:452)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1251/0x000840dcf840.run(Unknown
> Source)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.lambda$runInLeaderEventThread$3(DefaultLeaderElectionService.java:549)
> - locked <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1075/0x000840c23040.run(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture$AsyncRun.run(java.base@11.0.22/CompletableFuture.java:1736)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(java.base@11.0.22/ThreadPoolExecutor.java:1128)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(java.base@11.0.22/ThreadPoolExecutor.java:628)
> at java.lang.Thread.run(java.base@11.0.22/Thread.java:829)
> {code}
> {code}
> "jobmanager-io-thread-1" #636 daemon prio=5 os_prio=0 cpu=125.56ms
> elapsed=78699.01s tid=0x7f5321c6e800 nid=0x396 waiting for monitor entry
> [0x7f530567d000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.hasLeadership(DefaultLeaderElectionService.java:366)
> - waiting to lock <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElection.hasLeadership(DefaultLeaderElection.java:52)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.isValidLeader(JobMasterServiceLeadershipRunner.java:509)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.lambda$forwardIfValidLeader$15(JobMasterServiceLeadershipRunner.java:520)
> - locked <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner$$Lambda$1320/0x000840e1a840.accept(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture.uniWhenComplete(java.base@11.0.22/CompletableFuture.java:859)
> at
> java.util.concurrent.CompletableFuture$UniWhenComplete.tryFire(java.base@11.0.22/CompletableFuture.java:837)
> at
> java.uti
[jira] [Commented] (FLINK-34672) HA deadlock between JobMasterServiceLeadershipRunner and DefaultLeaderElectionService
[
https://issues.apache.org/jira/browse/FLINK-34672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17848648#comment-17848648
]
Matthias Pohl commented on FLINK-34672:
---
I'm still trying to find a reviewer. It's on my plate. But it's not a blocker
because the issue already existed in older versions of Flink:
{quote}
I also verified that this is not something that was introduced in Flink 1.18
with the FLIP-285 changes. AFAIS, it can also happen in 1.17- (I didn't check
the pre-FLINK-24038 code but only looked into release-1.17).
{quote}
> HA deadlock between JobMasterServiceLeadershipRunner and
> DefaultLeaderElectionService
> -
>
> Key: FLINK-34672
> URL: https://issues.apache.org/jira/browse/FLINK-34672
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
>Affects Versions: 1.17.2, 1.19.0, 1.18.1, 1.20.0
>Reporter: Chesnay Schepler
>Assignee: Matthias Pohl
>Priority: Major
> Labels: pull-request-available
> Fix For: 1.18.2, 1.20.0, 1.19.1
>
>
> We recently observed a deadlock in the JM within the HA system.
> (see below for the thread dump)
> [~mapohl] and I looked a bit into it and there appears to be a race condition
> when leadership is revoked while a JobMaster is being started.
> It appears to be caused by
> {{JobMasterServiceLeadershipRunner#createNewJobMasterServiceProcess}}
> forwarding futures while holding a lock; depending on whether the forwarded
> future is already complete the next stage may or may not run while holding
> that same lock.
> We haven't determined yet whether we should be holding that lock or not.
> {code}
> "DefaultLeaderElectionService-leadershipOperationExecutor-thread-1" #131
> daemon prio=5 os_prio=0 cpu=157.44ms elapsed=78749.65s tid=0x7f531f43d000
> nid=0x19d waiting for monitor entry [0x7f53084fd000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.runIfStateRunning(JobMasterServiceLeadershipRunner.java:462)
> - waiting to lock <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.revokeLeadership(JobMasterServiceLeadershipRunner.java:397)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.notifyLeaderContenderOfLeadershipLoss(DefaultLeaderElectionService.java:484)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1252/0x000840ddec40.accept(Unknown
> Source)
> at java.util.HashMap.forEach(java.base@11.0.22/HashMap.java:1337)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.onRevokeLeadershipInternal(DefaultLeaderElectionService.java:452)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1251/0x000840dcf840.run(Unknown
> Source)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.lambda$runInLeaderEventThread$3(DefaultLeaderElectionService.java:549)
> - locked <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1075/0x000840c23040.run(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture$AsyncRun.run(java.base@11.0.22/CompletableFuture.java:1736)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(java.base@11.0.22/ThreadPoolExecutor.java:1128)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(java.base@11.0.22/ThreadPoolExecutor.java:628)
> at java.lang.Thread.run(java.base@11.0.22/Thread.java:829)
> {code}
> {code}
> "jobmanager-io-thread-1" #636 daemon prio=5 os_prio=0 cpu=125.56ms
> elapsed=78699.01s tid=0x7f5321c6e800 nid=0x396 waiting for monitor entry
> [0x7f530567d000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.hasLeadership(DefaultLeaderElectionService.java:366)
> - waiting to lock <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElection.hasLeadership(DefaultLeaderElection.java:52)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.isValidLeader(JobMasterServiceLeadershipRunner.java:509)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.lambda$forwardIfValidLeader$15(JobMasterServiceLeadershipRunner.java:520)
> - locked <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner$$Lambda$1320
[jira] [Commented] (FLINK-34672) HA deadlock between JobMasterServiceLeadershipRunner and DefaultLeaderElectionService
[
https://issues.apache.org/jira/browse/FLINK-34672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17848646#comment-17848646
]
Hong Liang Teoh commented on FLINK-34672:
-
[~mapohl] do we know if this should be a blocker to the Flink 1.19.1 patch
release?
> HA deadlock between JobMasterServiceLeadershipRunner and
> DefaultLeaderElectionService
> -
>
> Key: FLINK-34672
> URL: https://issues.apache.org/jira/browse/FLINK-34672
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
>Affects Versions: 1.17.2, 1.19.0, 1.18.1, 1.20.0
>Reporter: Chesnay Schepler
>Assignee: Matthias Pohl
>Priority: Major
> Labels: pull-request-available
> Fix For: 1.18.2, 1.20.0, 1.19.1
>
>
> We recently observed a deadlock in the JM within the HA system.
> (see below for the thread dump)
> [~mapohl] and I looked a bit into it and there appears to be a race condition
> when leadership is revoked while a JobMaster is being started.
> It appears to be caused by
> {{JobMasterServiceLeadershipRunner#createNewJobMasterServiceProcess}}
> forwarding futures while holding a lock; depending on whether the forwarded
> future is already complete the next stage may or may not run while holding
> that same lock.
> We haven't determined yet whether we should be holding that lock or not.
> {code}
> "DefaultLeaderElectionService-leadershipOperationExecutor-thread-1" #131
> daemon prio=5 os_prio=0 cpu=157.44ms elapsed=78749.65s tid=0x7f531f43d000
> nid=0x19d waiting for monitor entry [0x7f53084fd000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.runIfStateRunning(JobMasterServiceLeadershipRunner.java:462)
> - waiting to lock <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.revokeLeadership(JobMasterServiceLeadershipRunner.java:397)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.notifyLeaderContenderOfLeadershipLoss(DefaultLeaderElectionService.java:484)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1252/0x000840ddec40.accept(Unknown
> Source)
> at java.util.HashMap.forEach(java.base@11.0.22/HashMap.java:1337)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.onRevokeLeadershipInternal(DefaultLeaderElectionService.java:452)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1251/0x000840dcf840.run(Unknown
> Source)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.lambda$runInLeaderEventThread$3(DefaultLeaderElectionService.java:549)
> - locked <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1075/0x000840c23040.run(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture$AsyncRun.run(java.base@11.0.22/CompletableFuture.java:1736)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(java.base@11.0.22/ThreadPoolExecutor.java:1128)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(java.base@11.0.22/ThreadPoolExecutor.java:628)
> at java.lang.Thread.run(java.base@11.0.22/Thread.java:829)
> {code}
> {code}
> "jobmanager-io-thread-1" #636 daemon prio=5 os_prio=0 cpu=125.56ms
> elapsed=78699.01s tid=0x7f5321c6e800 nid=0x396 waiting for monitor entry
> [0x7f530567d000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.hasLeadership(DefaultLeaderElectionService.java:366)
> - waiting to lock <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElection.hasLeadership(DefaultLeaderElection.java:52)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.isValidLeader(JobMasterServiceLeadershipRunner.java:509)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.lambda$forwardIfValidLeader$15(JobMasterServiceLeadershipRunner.java:520)
> - locked <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner$$Lambda$1320/0x000840e1a840.accept(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture.uniWhenComplete(java.base@11.0.22/CompletableFuture.java:859)
> at
> java.util.concurrent.CompletableFuture$UniWhenComplete.tryFire(java.base@11.0.22/CompletableFuture.java:837
[jira] [Commented] (FLINK-34672) HA deadlock between JobMasterServiceLeadershipRunner and DefaultLeaderElectionService
[
https://issues.apache.org/jira/browse/FLINK-34672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17828221#comment-17828221
]
Matthias Pohl commented on FLINK-34672:
---
IMHO, the both locks are still necessary because we're accessing the state in
the {{JobMasterServiceLeadershipRunner}} and the leadership in
{{DefaultLeaderElectionService}}. I also verified that this is not something
that was introduced in Flink 1.18 with the
[FLIP-285|https://cwiki.apache.org/confluence/display/FLINK/FLIP-285%3A+Refactoring+LeaderElection+to+make+Flink+support+multi-component+leader+election+out-of-the-box]
changes. AFAIS, it can also happen in 1.17- (I didn't check the
pre-FLINK-24038 code but only looked into {{release-1.17}}).
One solution would be to move the async callback of
[JobMasterServiceLeadershipRunner#forwardIfValidLeader|https://github.com/apache/flink/blob/c9fcb0c74b1354f4f0f1b7c7f62191b8cc6b5725/flink-runtime/src/main/java/org/apache/flink/runtime/jobmaster/JobMasterServiceLeadershipRunner.java#L547]
into the leader-operation executor of the {{DefaultLeaderElectionService}} to
force sequential execution of leadership-related operations.
This is only an issue in the {{JobMasterServiceLeadershipRunner}} because we're
executing the creation asynchronously in an io thread. The other place where we
check within the contender whether leadership is acquired is the
{{DefaultDispatcherRunner}}. But we're not doing any async calls there during
leadership handling (the {{DefaultDispatcherRunner}} is created directly in the
leader-operation executor while handling the leadership acquired event).
> HA deadlock between JobMasterServiceLeadershipRunner and
> DefaultLeaderElectionService
> -
>
> Key: FLINK-34672
> URL: https://issues.apache.org/jira/browse/FLINK-34672
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
>Affects Versions: 1.18.1
>Reporter: Chesnay Schepler
>Priority: Major
> Fix For: 1.18.2, 1.20.0, 1.19.1
>
>
> We recently observed a deadlock in the JM within the HA system.
> (see below for the thread dump)
> [~mapohl] and I looked a bit into it and there appears to be a race condition
> when leadership is revoked while a JobMaster is being started.
> It appears to be caused by
> {{JobMasterServiceLeadershipRunner#createNewJobMasterServiceProcess}}
> forwarding futures while holding a lock; depending on whether the forwarded
> future is already complete the next stage may or may not run while holding
> that same lock.
> We haven't determined yet whether we should be holding that lock or not.
> {code}
> "DefaultLeaderElectionService-leadershipOperationExecutor-thread-1" #131
> daemon prio=5 os_prio=0 cpu=157.44ms elapsed=78749.65s tid=0x7f531f43d000
> nid=0x19d waiting for monitor entry [0x7f53084fd000]
>java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.runIfStateRunning(JobMasterServiceLeadershipRunner.java:462)
> - waiting to lock <0xf1c0e088> (a java.lang.Object)
> at
> org.apache.flink.runtime.jobmaster.JobMasterServiceLeadershipRunner.revokeLeadership(JobMasterServiceLeadershipRunner.java:397)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.notifyLeaderContenderOfLeadershipLoss(DefaultLeaderElectionService.java:484)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1252/0x000840ddec40.accept(Unknown
> Source)
> at java.util.HashMap.forEach(java.base@11.0.22/HashMap.java:1337)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.onRevokeLeadershipInternal(DefaultLeaderElectionService.java:452)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1251/0x000840dcf840.run(Unknown
> Source)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService.lambda$runInLeaderEventThread$3(DefaultLeaderElectionService.java:549)
> - locked <0xf0e3f4d8> (a java.lang.Object)
> at
> org.apache.flink.runtime.leaderelection.DefaultLeaderElectionService$$Lambda$1075/0x000840c23040.run(Unknown
> Source)
> at
> java.util.concurrent.CompletableFuture$AsyncRun.run(java.base@11.0.22/CompletableFuture.java:1736)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(java.base@11.0.22/ThreadPoolExecutor.java:1128)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(java.base@11.0.22/ThreadPoolExecutor.java:628)
> at java.lang.Thread.run(java.base@11.0.22/Thread.java:829)
> {code}
> {code}
> "jobmanager-io-thread-1" #636 daemon pri
