Re: Flink YARN app terminated before the client receives the result

Mon, 30 Mar 2020 02:49:01 -0700

I think we have to take a step back here. For per-job (YARN) mode, the general problem is that there are two systems that can do shutdown (and other things) and two clients. There is YARN and there is Flink, and Flink is YARN inside YARN, in a way. The solution, I think, is that cancellation for YARN mode should go though YARN, not through Flink. Then there can be no races or other issues with the cluster shutting down before it has a chance to send a response.
Btw, the same goes for "attached mode" where a client waits for job completion. IMO, this should also go through YARN and not the Flink REST client. 

What do you think?

Best,
Aljoscha

On 20.03.20 15:15, Till Rohrmann wrote:

Yes you are right that `thenAcceptAsync` only breaks the control flow but
it does not guarantee that the `RestServer` has actually sent the response
to the client. Maybe we also need something similar to FLINK-10309 [1]. The
problem I see with this approach is that it makes all RestHandlers stateful.

[1] https://issues.apache.org/jira/browse/FLINK-10309

Cheers,
Till

On Fri, Mar 20, 2020 at 2:26 PM DONG, Weike <kyled...@connect.hku.hk> wrote:


Hi Tison & Till,

Changing *thenAccept *into *thenAcceptAsync *in the
MiniDispatcher#cancelJob does not help to solve the problem in my
environment. However, I have found that adding a* Thread.sleep(2000) *before
the return of JobCancellationHandler#handleRequest solved the problem (at
least the symptom goes away). As this is only a dirty hack, I will try to
get a more decent solution to this problem.

Sincerely,
Weike

On Tue, Mar 17, 2020 at 11:11 PM tison <wander4...@gmail.com> wrote:


JIRA created as https://jira.apache.org/jira/browse/FLINK-16637

Best,
tison.


Till Rohrmann <trohrm...@apache.org> 于2020年3月17日周二 下午5:57写道：


  @Tison could you create an issue to track the problem. Please also link
the uploaded log file for further debugging.

I think the reason why it worked in Flink 1.9 could have been that we
had a async callback in the longer chain which broke the flow of execution
and allowed to send the response. This is no longer the case. As an easy
fix one could change thenAccept into thenAcceptAsync in the
MiniDispatcher#cancelJob. But this only solves symptoms. Maybe we should
think about allowing not only StatusHandler to close asynchronously. At the
moment we say that all other handler shut down immediately (see
AbstractHandler#closeHandlerAsync). But the problem with this change would
be that all handler would become stateful because they would need to
remember whether a request is currently ongoing or not.

Cheers,
Till

On Tue, Mar 17, 2020 at 9:01 AM DONG, Weike <kyled...@connect.hku.hk>
wrote:


Hi Tison & Till and all,

I have uploaded the client, taskmanager and jobmanager log to Gist (
https://gist.github.com/kylemeow/500b6567368316ec6f5b8f99b469a49f),
and I
can reproduce this bug every time when trying to cancel Flink 1.10 jobs
on
YARN.

Besides, in earlier Flink versions like 1.9, the REST API for
*cancelling
job with a savepoint *sometimes throws exceptions to the client side
due to
early shutdown of the server, even though the savepoint was successfully
completed by reviewing the log, however when using the newly introduced
*stop* API, that bug disappeared, however, *cancel* API seems to be
buggy
now.

Best,
Weike

On Tue, Mar 17, 2020 at 10:17 AM tison <wander4...@gmail.com> wrote:


edit: previously after the cancellation we have a longer call chain to
#jobReachedGloballyTerminalState which does the archive job & JM

graceful

showdown, which might take some time so that ...

Best,
tison.


tison <wander4...@gmail.com> 于2020年3月17日周二 上午10:13写道：


Hi Weike & Till,

I agree with Till and it is also the analysis from my side. However,

it

seems even if we don't have FLINK-15116, it is still possible that we
complete the cancel future but the cluster got shutdown before it

properly

delivered the response.

There is one thing strange that this behavior almost reproducible, it
should be a possible order but not always. Maybe previous we have to
firstly cancel the job which has a long call chain so that it

happens we

have enough time to delivered the response.

But the resolution looks like we introduce some
synchronization/finalization logics that clear these outstanding

future

with best effort before the cluster(RestServer) down.

Best,
tison.


Till Rohrmann <trohrm...@apache.org> 于2020年3月17日周二 上午4:12写道：


Hi Weike,

could you share the complete logs with us? Attachments are being
filtered out by the Apache mail server but it works if you upload

the logs

somewhere (e.g. https://gist.github.com/) and then share the link

with

us. Ideally you run the cluster with DEBUG log settings.

I assume that you are running Flink 1.10, right?

My suspicion is that this behaviour has been introduced with

FLINK-15116

[1]. It looks as if we complete the shutdown future in
MiniDispatcher#cancelJob before we return the response to the
RestClusterClient. My guess is that this triggers the shutdown of

the

RestServer which then is not able to serve the response to the

client. I'm

pulling in Aljoscha and Tison who introduced this change. They

might be

able to verify my theory and propose a solution for it.

[1] https://issues.apache.org/jira/browse/FLINK-15116

Cheers,
Till

On Fri, Mar 13, 2020 at 7:50 AM DONG, Weike <

kyled...@connect.hku.hk>

wrote:


Hi Yangze and all,

I have tried numerous times, and this behavior persists.

Below is the tail log of taskmanager.log:

2020-03-13 12:06:14.240 [flink-akka.actor.default-dispatcher-3]

INFO

  org.apache.flink.runtime.taskexecutor.slot.TaskSlotTableImpl  -

Free slot

TaskSlot(index:0, state:ACTIVE, resource profile:
ResourceProfile{cpuCores=1.0000000000000000, taskHeapMemory=1.503gb
(1613968148 bytes), taskOffHeapMemory=0 bytes,

managedMemory=1.403gb

(1505922928 bytes), networkMemory=359.040mb (376480732 bytes)},
allocationId: d3acaeac3db62454742e800b5410adfd, jobId:
d0a674795be98bd2574d9ea3286801cb).
2020-03-13 12:06:14.244 [flink-akka.actor.default-dispatcher-3]

INFO

  org.apache.flink.runtime.taskexecutor.JobLeaderService  - Remove

job

d0a674795be98bd2574d9ea3286801cb from job leader monitoring.
2020-03-13 12:06:14.244 [flink-akka.actor.default-dispatcher-3]

INFO

  org.apache.flink.runtime.taskexecutor.TaskExecutor  - Close

JobManager

connection for job d0a674795be98bd2574d9ea3286801cb.
2020-03-13 12:06:14.250 [flink-akka.actor.default-dispatcher-3]

INFO

  org.apache.flink.runtime.taskexecutor.TaskExecutor  - Close

JobManager

connection for job d0a674795be98bd2574d9ea3286801cb.
2020-03-13 12:06:14.250 [flink-akka.actor.default-dispatcher-3]

INFO

  org.apache.flink.runtime.taskexecutor.JobLeaderService  - Cannot

reconnect

to job d0a674795be98bd2574d9ea3286801cb because it is not

registered.

2020-03-13 12:06:19.744 [SIGTERM handler] INFO
  org.apache.flink.yarn.YarnTaskExecutorRunner  - RECEIVED SIGNAL

15:

SIGTERM. Shutting down as requested.
2020-03-13 12:06:19.744 [SIGTERM handler] INFO
  org.apache.flink.yarn.YarnTaskExecutorRunner  - RECEIVED SIGNAL

15:

SIGTERM. Shutting down as requested.
2020-03-13 12:06:19.745 [PermanentBlobCache shutdown hook] INFO
  org.apache.flink.runtime.blob.PermanentBlobCache  - Shutting down

BLOB

cache
2020-03-13 12:06:19.749 [FileChannelManagerImpl-netty-shuffle

shutdown

hook] INFO  org.apache.flink.runtime.io.disk.FileChannelManagerImpl

-

FileChannelManager removed spill file directory


/data/emr/yarn/local/usercache/hadoop/appcache/application_1562207369540_0135/flink-netty-shuffle-65cd4ebb-51f4-48a9-8e3c-43e431bca46d

2020-03-13 12:06:19.750 [TaskExecutorLocalStateStoresManager

shutdown

hook] INFO


org.apache.flink.runtime.state.TaskExecutorLocalStateStoresManager  -

Shutting down TaskExecutorLocalStateStoresManager.
2020-03-13 12:06:19.750 [TransientBlobCache shutdown hook] INFO
  org.apache.flink.runtime.blob.TransientBlobCache  - Shutting down

BLOB

cache
2020-03-13 12:06:19.751 [FileChannelManagerImpl-io shutdown hook]

INFO

  org.apache.flink.runtime.io.disk.FileChannelManagerImpl  -
FileChannelManager removed spill file directory


/data/emr/yarn/local/usercache/hadoop/appcache/application_1562207369540_0135/flink-io-67ad5c3a-aec6-42be-ab1f-0ce3841fc4bd

2020-03-13 12:06:19.752 [FileCache shutdown hook] INFO
  org.apache.flink.runtime.filecache.FileCache  - removed file cache
directory


/data/emr/yarn/local/usercache/hadoop/appcache/application_1562207369540_0135/flink-dist-cache-65075ee3-e009-4978-a9d8-ec010e6f4b31


As the tail log of jobmanager.log is kind of lengthy, I have

attached

it in this mail.

 From what I have seen, the TaskManager and JobManager shut down by
themselves, however, I have noticed some Netty exceptions (from

the stack

trace, it is part of the REST handler) like:

ERROR


org.apache.flink.shaded.netty4.io.netty.util.concurrent.DefaultPromise.rejectedExecution

  - Failed to submit a listener notification task. Event loop shut

down?

java.util.concurrent.RejectedExecutionException: event executor
terminated

Thus I suppose that these exceptions might be the actual cause of
premature termination of the REST server, and I am still looking

into the

real cause of this.

Best,
Weike

On Fri, Mar 13, 2020 at 1:45 PM Yangze Guo <karma...@gmail.com>

wrote:



Would you mind to share more information about why the task

executor

is killed? If it is killed by Yarn, you might get such info in

Yarn

NM/RM logs.

Best,
Yangze Guo

Best,
Yangze Guo


On Fri, Mar 13, 2020 at 12:31 PM DONG, Weike <

kyled...@connect.hku.hk>

wrote:


Hi,

Recently I have encountered a strange behavior of Flink on YARN,

which is that when I try to cancel a Flink job running in per-job

mode on

YARN using commands like


"cancel -m yarn-cluster -yid application_1559388106022_9412

ed7e2e0ab0a7316c1b65df6047bc6aae"


the client happily found and connected to ResourceManager and

then

stucks at

Found Web Interface 172.28.28.3:50099 of application

'application_1559388106022_9412'.


And after one minute, an exception is thrown at the client side:
Caused by: org.apache.flink.util.FlinkException: Could not

cancel

job ed7e2e0ab0a7316c1b65df6047bc6aae.

     at



org.apache.flink.client.cli.CliFrontend.lambda$cancel$7(CliFrontend.java:545)

     at



org.apache.flink.client.cli.CliFrontend.runClusterAction(CliFrontend.java:843)

     at



org.apache.flink.client.cli.CliFrontend.cancel(CliFrontend.java:538)

     at



org.apache.flink.client.cli.CliFrontend.parseParametersWithException(CliFrontend.java:917)

     at



org.apache.flink.client.cli.CliFrontend.lambda$mainWithReturnCodeAndException$10(CliFrontend.java:988)

     at java.security.AccessController.doPrivileged(Native

Method)

     at javax.security.auth.Subject.doAs(Subject.java:422)
     at



org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1754)

     ... 20 more
Caused by: java.util.concurrent.TimeoutException
     at



java.util.concurrent.CompletableFuture.timedGet(CompletableFuture.java:1771)

     at



java.util.concurrent.CompletableFuture.get(CompletableFuture.java:1915)

     at



org.apache.flink.client.cli.CliFrontend.lambda$cancel$7(CliFrontend.java:543)

     ... 27 more

Then I discovered that the YARN app has already terminated with

FINISHED state and KILLED final status, like below.


And after digging into the log of this finished YARN app, I have

found that TaskManager had already received the SIGTERM signal and
terminated gracefully.

org.apache.flink.yarn.YarnTaskExecutorRunner  - RECEIVED SIGNAL

15:

SIGTERM. Shutting down as requested.


Also, the log of JobManager shows that it terminated with exit

code

0.

Terminating cluster entrypoint process YarnJobClusterEntrypoint

with

exit code 0


However, the JobManager did not return anything to the client

before

its shutdown, which is different from previous versions (like

Flink 1.9).


I wonder if this is a new bug on the flink-clients or flink-yarn

module?


Thank you : )

Sincerely,
Weike



































					Reply via email to