It is a very useful optimization in large scale development. In blink we use a thread pool to launch container, and a flink cluster with more than 5000 Yarn containers could work well. Hope Xintong Song could contribute it to flink soon.

Xintong Song <tonysong820@gmail.com> 于2019年7月10日周三 下午2:20写道:
Thanks for the kindly offer, Qi. 

I think this work should not take much time, so I can take care of it. It's just the community is currently under feature freeze for release 1.9, so we need to wait until the release code branch being cut.


Thank you~

Xintong Song



On Wed, Jul 10, 2019 at 1:55 PM qi luo <luoqi.bd@gmail.com> wrote:
Thanks Xintong and Haibo, I’ve found the fix in the blink branch. 

We’re also glad to help contribute this patch to community version, in case you don’t have time.

Regards,
Qi

On Jul 10, 2019, at 11:51 AM, Haibo Sun <sunhaibotb@163.com> wrote:

Hi, Qi

Sorry, by talking to Xintong Song offline, I made sure that some of what I said before is wrong. Please refer to Xintong's answer.

Best,
Haibo

At 2019-07-10 11:37:16, "Xintong Song" <tonysong820@gmail.com> wrote:
Hi Qi,

Thanks for reporting this problem. I think you are right. Starting large amount of TMs in main thread on YARN could take relative long time, causing RM to become unresponsive. In our enterprise version Blink, we actually have a thread pool for starting TMs. I think we should contribute this feature to the community version as well. I created a JIRA ticket from which you can track the progress of this issue. 

For solving your problem at the moment, I agree with Haibo that configure a larger registration timeout could be a workaround.

Thank you~
Xintong Song



On Wed, Jul 10, 2019 at 10:37 AM Haibo Sun <sunhaibotb@163.com> wrote:
Hi, Qi

According to our experience, it is no problem to allocate more than 1000 containers when the registration timeout is set 5 minutes . Perhaps there are other reasons? Or you can try to increase the value of `taskmanager.registration.timeout`. For allocating containers using multi-thread, I personally think it's going to get very complicated, and the more recommended way is to put some waiting works into asynchronous processing, so as to liberate the main thread.

Best,
Haibo


At 2019-07-09 21:05:51, "qi luo" <luoqi.bd@gmail.com> wrote:
Hi guys,

We’re using latest version Flink YarnResourceManager, but our job startup occasionally hangs when allocating many Yarn containers (e.g. >1000). I checked the related code in YarnResourceManager as below:

<PastedGraphic-1.tiff>

It seems that it handles all allocated containers and starts TM in main thread. Thus when containers allocations are heavy, the RM thread becomes unresponsive (such as no response to TM heartbeats, see TM logs as below).

Any idea on how to better handle such case (e.g. multi-threading to handle allocated containers) would be very appreciated. Thanks!

Regards,
Qi 


———————————————————————— 
TM log:

2019-07-09 13:56:59,110 INFO  org.apache.flink.runtime.taskexecutor.TaskExecutor            - Connecting to ResourceManager akka.tcp://flink@xxx/user/resourcemanager(00000000000000000000000000000000).
2019-07-09 14:00:01,138 INFO  org.apache.flink.runtime.taskexecutor.TaskExecutor            - Could not resolve ResourceManager address akka.tcp://flink@xxx/user/resourcemanager, retrying in 10000 ms: Ask timed out on [ActorSelection[Anchor(akka.tcp://flink@xxx/), Path(/user/resourcemanager)]] after [182000 ms]. Sender[null] sent message of type "akka.actor.Identify"..
2019-07-09 14:01:59,137 ERROR org.apache.flink.runtime.taskexecutor.TaskExecutor            - Fatal error occurred in TaskExecutor akka.tcp://flink@xxx/user/taskmanager_0.
org.apache.flink.runtime.taskexecutor.exceptions.RegistrationTimeoutException: Could not register at the ResourceManager within the specified maximum registration duration 300000 ms. This indicates a problem with this instance. Terminating now.
	at org.apache.flink.runtime.taskexecutor.TaskExecutor.registrationTimeout(TaskExecutor.java:1023)
	at org.apache.flink.runtime.taskexecutor.TaskExecutor.lambda$startRegistrationTimeout$3(TaskExecutor.java:1009)
	at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.handleRunAsync(AkkaRpcActor.java:332)
	at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.handleRpcMessage(AkkaRpcActor.java:158)
	at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.onReceive(AkkaRpcActor.java:142)
	at akka.actor.UntypedActor$$anonfun$receive$1.applyOrElse(UntypedActor.scala:165)
	at akka.actor.Actor$class.aroundReceive(Actor.scala:502)
	at akka.actor.UntypedActor.aroundReceive(UntypedActor.scala:95)
	at akka.actor.ActorCell.receiveMessage(ActorCell.scala:526)
	at akka.actor.ActorCell.invoke(ActorCell.scala:495)
	at akka.dispatch.Mailbox.processMailbox(Mailbox.scala:257)
	at akka.dispatch.Mailbox.run(Mailbox.scala:224)
	at akka.dispatch.Mailbox.exec(Mailbox.scala:234)
	at scala.concurrent.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260)
	at scala.concurrent.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339)
	at scala.concurrent.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979)
	at scala.concurrent.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107)