Hi, Unfortunately the thing I described does indeed happen every time. As mentioned in the first email, I am running on Kubernetes so certain things could be different compared to just a standalone cluster.
Any ideas for workarounds are welcome, as this problem basically prevents me from using HA. Thanks, Encho On Wed, Aug 29, 2018 at 5:15 AM vino yang <yanghua1...@gmail.com> wrote: > Hi Encho, > > From your description, I feel that there are extra bugs. > > About your description: > > *- Start both job managers* > *- Start a batch job in JobManager 1 and let it finish* > *The jobgraphs in both Zookeeper and HDFS remained.* > > Is it necessarily happening every time? > > In the Standalone cluster, the problems we encountered were sporadic. > > Thanks, vino. > > Encho Mishinev <encho.mishi...@gmail.com> 于2018年8月28日周二 下午8:07写道: > >> Hello Till, >> >> I spend a few more hours testing and looking at the logs and it seems >> like there's a more general problem here. While the two job managers are >> active neither of them can properly delete jobgraphs. The above problem I >> described comes from the fact that Kubernetes gets JobManager 1 quickly >> after I manually kill it, so when I stop the job on JobManager 2 both are >> alive. >> >> I did a very simple test: >> >> - Start both job managers >> - Start a batch job in JobManager 1 and let it finish >> The jobgraphs in both Zookeeper and HDFS remained. >> >> On the other hand if we do: >> >> - Start only JobManager 1 (again in HA mode) >> - Start a batch job and let it finish >> The jobgraphs in both Zookeeper and HDFS are deleted fine. >> >> It seems like the standby manager still leaves some kind of lock on the >> jobgraphs. Do you think that's possible? Have you seen a similar problem? >> The only logs that appear on the standby manager while waiting are of the >> type: >> >> 2018-08-28 11:54:10,789 INFO >> org.apache.flink.runtime.jobmanager.ZooKeeperSubmittedJobGraphStore - >> Recovered SubmittedJobGraph(9e0a109b57511930c95d3b54574a66e3, null). >> >> Note that this log appears on the standby jobmanager immediately when a >> new job is submitted to the active jobmanager. >> Also note that the blobs and checkpoints are cleared fine. The problem is >> only for jobgraphs both in ZooKeeper and HDFS. >> >> Trying to access the UI of the standby manager redirects to the active >> one, so it is not a problem of them not knowing who the leader is. Do you >> have any ideas? >> >> Thanks a lot, >> Encho >> >> On Tue, Aug 28, 2018 at 10:27 AM Till Rohrmann <trohrm...@apache.org> >> wrote: >> >>> Hi Encho, >>> >>> thanks a lot for reporting this issue. The problem arises whenever the >>> old leader maintains the connection to ZooKeeper. If this is the case, then >>> ephemeral nodes which we create to protect against faulty delete operations >>> are not removed and consequently the new leader is not able to delete the >>> persisted job graph. So one thing to check is whether the old JM still has >>> an open connection to ZooKeeper. The next thing to check is the session >>> timeout of your ZooKeeper cluster. If you stop the job within the session >>> timeout, then it is also not guaranteed that ZooKeeper has detected that >>> the ephemeral nodes of the old JM must be deleted. In order to understand >>> this better it would be helpful if you could tell us the timing of the >>> different actions. >>> >>> Cheers, >>> Till >>> >>> On Tue, Aug 28, 2018 at 8:17 AM vino yang <yanghua1...@gmail.com> wrote: >>> >>>> Hi Encho, >>>> >>>> A temporary solution can be used to determine if it has been cleaned up >>>> by monitoring the specific JobID under Zookeeper's "/jobgraph". >>>> Another solution, modify the source code, rudely modify the cleanup >>>> mode to the synchronous form, but the flink operation Zookeeper's path >>>> needs to obtain the corresponding lock, so it is dangerous to do so, and it >>>> is not recommended. >>>> I think maybe this problem can be solved in the next version. It >>>> depends on Till. >>>> >>>> Thanks, vino. >>>> >>>> Encho Mishinev <encho.mishi...@gmail.com> 于2018年8月28日周二 下午1:17写道: >>>> >>>>> Thank you very much for the info! Will keep track of the progress. >>>>> >>>>> In the meantime is there any viable workaround? It seems like HA >>>>> doesn't really work due to this bug. >>>>> >>>>> On Tue, Aug 28, 2018 at 4:52 AM vino yang <yanghua1...@gmail.com> >>>>> wrote: >>>>> >>>>>> About some implementation mechanisms. >>>>>> Flink uses Zookeeper to store JobGraph (Job's description information >>>>>> and metadata) as a basis for Job recovery. >>>>>> However, previous implementations may cause this information to not >>>>>> be properly cleaned up because it is asynchronously deleted by a >>>>>> background >>>>>> thread. >>>>>> >>>>>> Thanks, vino. >>>>>> >>>>>> vino yang <yanghua1...@gmail.com> 于2018年8月28日周二 上午9:49写道: >>>>>> >>>>>>> Hi Encho, >>>>>>> >>>>>>> This is a problem already known to the Flink community, you can >>>>>>> track its progress through FLINK-10011[1], and currently Till is fixing >>>>>>> this issue. >>>>>>> >>>>>>> [1]: https://issues.apache.org/jira/browse/FLINK-10011 >>>>>>> >>>>>>> Thanks, vino. >>>>>>> >>>>>>> Encho Mishinev <encho.mishi...@gmail.com> 于2018年8月27日周一 下午10:13写道: >>>>>>> >>>>>>>> I am running Flink 1.5.3 with two job managers and two task >>>>>>>> managers in Kubernetes along with HDFS and Zookeeper in >>>>>>>> high-availability >>>>>>>> mode. >>>>>>>> >>>>>>>> My problem occurs after the following actions: >>>>>>>> - Upload a .jar file to jobmanager-1 >>>>>>>> - Run a streaming job from the jar on jobmanager-1 >>>>>>>> - Wait for 1 or 2 checkpoints to succeed >>>>>>>> - Kill pod of jobmanager-1 >>>>>>>> After a short delay, jobmanager-2 takes leadership and correctly >>>>>>>> restores the job and continues it >>>>>>>> - Stop job from jobmanager-2 >>>>>>>> >>>>>>>> At this point all seems well, but the problem is that jobmanager-2 >>>>>>>> does not clean up anything that was left from jobmanager-1. This means >>>>>>>> that >>>>>>>> both in HDFS and in Zookeeper remain job graphs, which later on >>>>>>>> obstruct >>>>>>>> any work of both managers as after any reset they unsuccessfully try to >>>>>>>> restore a non-existent job and fail over and over again. >>>>>>>> >>>>>>>> I am quite certain that jobmanager-2 does not know about any of >>>>>>>> jobmanager-1’s files since the Zookeeper logs reveal that it tries to >>>>>>>> duplicate job folders: >>>>>>>> >>>>>>>> 2018-08-27 13:11:00,038 [myid:] - INFO [ProcessThread(sid:0 >>>>>>>> cport:2181)::PrepRequestProcessor@648] - Got user-level >>>>>>>> KeeperException when processing sessionid:0x1657aa15e480033 type:create >>>>>>>> cxid:0x46 zxid:0x1ab txntype:-1 reqpath:n/a Error >>>>>>>> Path:/flink/default/jobgraphs/83bfa359ca59ce1d4635e18e16651e15/bbb259fd-7826-4950-bc7c-c2be23346c77 >>>>>>>> Error:KeeperErrorCode = NodeExists for >>>>>>>> /flink/default/jobgraphs/83bfa359ca59ce1d4635e18e16651e15/bbb259fd-7826-4950-bc7c-c2be23346c77 >>>>>>>> >>>>>>>> 2018-08-27 13:11:02,296 [myid:] - INFO [ProcessThread(sid:0 >>>>>>>> cport:2181)::PrepRequestProcessor@648] - Got user-level >>>>>>>> KeeperException when processing sessionid:0x1657aa15e480033 type:create >>>>>>>> cxid:0x5c zxid:0x1ac txntype:-1 reqpath:n/a Error >>>>>>>> Path:/flink/default/checkpoint-counter/83bfa359ca59ce1d4635e18e16651e15 >>>>>>>> Error:KeeperErrorCode = NodeExists for >>>>>>>> /flink/default/checkpoint-counter/83bfa359ca59ce1d4635e18e16651e15 >>>>>>>> >>>>>>>> Also jobmanager-2 attempts to delete the jobgraphs folder in >>>>>>>> Zookeeper when the job is stopped, but fails since there are leftover >>>>>>>> files >>>>>>>> in it from jobmanager-1: >>>>>>>> >>>>>>>> 2018-08-27 13:12:13,406 [myid:] - INFO [ProcessThread(sid:0 >>>>>>>> cport:2181)::PrepRequestProcessor@648] - Got user-level >>>>>>>> KeeperException when processing sessionid:0x1657aa15e480033 type:delete >>>>>>>> cxid:0xa8 zxid:0x1bd txntype:-1 reqpath:n/a Error >>>>>>>> Path:/flink/default/jobgraphs/83bfa359ca59ce1d4635e18e16651e15 >>>>>>>> Error:KeeperErrorCode = Directory not empty for >>>>>>>> /flink/default/jobgraphs/83bfa359ca59ce1d4635e18e16651e15 >>>>>>>> >>>>>>>> I’ve noticed that when restoring the job, it seems like >>>>>>>> jobmanager-2 does not get anything more than jobID, while it perhaps >>>>>>>> needs >>>>>>>> some metadata? Here is the log that seems suspicious to me: >>>>>>>> >>>>>>>> 2018-08-27 13:09:18,113 INFO >>>>>>>> org.apache.flink.runtime.jobmanager.ZooKeeperSubmittedJobGraphStore - >>>>>>>> Recovered SubmittedJobGraph(83bfa359ca59ce1d4635e18e16651e15, null). >>>>>>>> >>>>>>>> All other logs seem fine in jobmanager-2, it doesn’t seem to be >>>>>>>> aware that it’s overwriting anything or not deleting properly. >>>>>>>> >>>>>>>> My question is - what is the intended way for the job managers to >>>>>>>> correctly exchange metadata in HA mode and why is it not working for >>>>>>>> me? >>>>>>>> >>>>>>>> Thanks in advance! >>>>>>> >>>>>>>