[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: PartitionNotFoundException after deployment

Looks pretty clear that one operator takes too long to start (even on the UI it shows it in the created state for far too long). Any idea what might cause this delay? It actually often crashes on Akka ask timeout during scheduling the node.


Piotr Nowojski <piotr@xxxxxxxxxxxxxxxxx> ezt írta (időpont: 2018. máj. 4., P, 15:33):
Ufuk: I don’t know why.

+1 for your other suggestions.


> On 4 May 2018, at 14:52, Ufuk Celebi <ufuk@xxxxxxxxxxxxxxxxx> wrote:
> Hey Gyula!
> I'm including Piotr and Nico (cc'd) who have worked on the network
> stack in the last releases.
> Registering the network structures including the intermediate results
> actually happens **before** any state is restored. I'm not sure why
> this reproducibly happens when you restore state. @Nico, Piotr: any
> ideas here?
> In general I think what happens here is the following:
> - a task requests the result of a local upstream producer, but that
> one has not registered its intermediate result yet
> - this should result in a retry of the request with some backoff
> (controlled via the config params you mention
> taskmanager.network.request-backoff.max,
> taskmanager.network.request-backoff.initial)
> As a first step I would set logging to DEBUG and check the TM logs for
> messages like "Retriggering partition request {}:{}."
> You can also check the SingleInputGate code which has the logic for
> retriggering requests.
> – Ufuk
> On Fri, May 4, 2018 at 10:27 AM, Gyula Fóra <gyula.fora@xxxxxxxxx> wrote:
>> Hi Ufuk,
>> Do you have any quick idea what could cause this problems in flink 1.4.2?
>> Seems like one operator takes too long to deploy and downstream tasks error
>> out on partition not found. This only seems to happen when the job is
>> restored from state and in fact that operator has some keyed and operator
>> state as well.
>> Deploying the same job from empty state works well. We tried increasing the
>> taskmanager.network.request-backoff.max that didnt help.
>> It would be great if you have some pointers where to look further, I havent
>> seen this happening before.
>> Thank you!
>> Gyula
>> The errror:
>> org.apache.flink.runtime.io.network.partition.: Partition
>> 4c5e9cd5dd410331103f51127996068a@b35ef4ffe25e3d17c5d6051ebe2860cd not found.
>>    at
>> org.apache.flink.runtime.io.network.partition.ResultPartitionManager.createSubpartitionView(ResultPartitionManager.java:77)
>>    at
>> org.apache.flink.runtime.io.network.partition.consumer.LocalInputChannel.requestSubpartition(LocalInputChannel.java:115)
>>    at
>> org.apache.flink.runtime.io.network.partition.consumer.LocalInputChannel$1.run(LocalInputChannel.java:159)
>>    at java.util.TimerThread.mainLoop(Timer.java:555)
>>    at java.util.TimerThread.run(Timer.java:505)
> --
> Data Artisans GmbH | Stresemannstr. 121a | 10963 Berlin
> info@xxxxxxxxxxxxxxxxx
> +49-30-43208879
> Registered at Amtsgericht Charlottenburg - HRB 158244 B
> Managing Directors: Dr. Kostas Tzoumas, Dr. Stephan Ewen