Re: [External Sender] Debugging "Container is running beyond physical memory limits" on YARN for a long running streaming job

[Shubham Kumar <sh...@gmail.com>]

@Kye , Thanks for your suggestions, we are using one yarn app per job mode
and your point is still valid in Flink 1.10 as per docs, it does make sense
to avoid dynamic classloading for such jobs. Also, we seemed to have enough
off heap for resources mentioned and what turned out to be the issue was
RocksDB memory usage (check below).

@Xintong, Yeah, I did try out the solution, the problem is definitely due
to RocksDB, however the problem got solved by something else:

Short answer:
Setting this property in flink-conf.yaml solves the issue:

> state.backend.rocksdb.managed.memory : false


Long answer:

I observed that the OOM kills are a function of the number of restarts
rather than the time for which the application is running. For every
restart, the Taskmanager's RES memory rises by 3.5GB (which is the Flink
managed memory allotted to TM). So, it could only withstand 2-3 restarts
after which OOM kills become frequent as now the other TM will start
getting killed. I enabled RocksDB  block-cache usage metric and it rises up
until it reaches ~3.5 GB.

At this point I tried setting

> containerized.taskmanager.env.MALLOC_ARENA_MAX : 2


This did seem to reduce memory increase for few of the task managers(for
e.g. if there are 12 task managers,  after a restart the RES memory
increases by 3.5 GB for only few of them but not for others), but didn't
solve the issue for me and OOM kills begin to occur after 4-5 restarts. I
also tried setting it to 1, but got similar results. I didn't try using
jemalloc because as per the JIRA issue [1], MALLOC_ARENA_MAX solution
intends to produce similar results.

After setting state.backend.rocksdb.managed.memory: false, the TM RES
memory doesn't increase after any number of restarts, infact after enabling
RocksDB cache usage metrics, it shows around only ~100MB usage (ofcourse
its dependent on the operators and state involved in the job). This might
indicate that Flink is trying to allot more memory than required for
RocksDB and also upon restart the RES memory rises again which is
definitely not the intended behavior.

[1]: https://issues.apache.org/jira/browse/FLINK-18712

Thanks
Shubham


On Fri, Sep 25, 2020 at 8:46 PM Kye Bae <ky...@capitalone.com> wrote:

> Not sure about Flink 1.10.x. Can share a few things up to Flink 1.9.x:
>
> 1. If your Flink cluster runs only one job, avoid using dynamic
> classloader for your job: start it from one of the Flink class paths. As of
> Flink 1.9.x, using the dynamic classloader results in the same classes
> getting loaded every time the job restarts (self-recovery or otherwise),
> and it could eat up all the JVM "off-heap" memory. Yarn seems to
> immediately kill the container when that happens.
>
> 2. Be sure to leave enough for the JVM "off-heap" area: GC + code cache +
> thread stacks + other Java internal resources end up there.
>
> -K
>
> On Sat, Sep 19, 2020 at 12:09 PM Shubham Kumar <sh...@gmail.com>
> wrote:
>
>> Hey everyone,
>>
>> We had deployed a streaming job using Flink 1.10.1 one month back and now
>> we are encountering a Yarn container killed due to memory issues very
>> frequently. I am trying to figure out the root cause of this issue in order
>> to fix it.
>>
>> We have a streaming job whose basic structure looks like this:
>> - Read 6 kafka streams and combine stats from them (union) to form a
>> single stream
>> - stream.keyBy(MyKey)
>>              .window(TumblingEventTimeWindows.of(Time.minutes(1)))
>>              .reduce(MyReduceFunction)
>>              .addSink(new FlinkKafkaProducer011<>...);
>>
>> We are using RocksDB as state backend. In flink-conf.yaml, we used
>> taskmanager.memory.process.size = 10GB with a parallelism of 12 and only
>> one slot per task manager.
>>
>> So, a taskmanager process gets started with the following memory
>> components as indicated in logs:
>>
>> TaskExecutor container... will be started on ... with
>>> TaskExecutorProcessSpec {cpuCores=1.0, frameworkHeapSize=128.000mb (
>>> 134217728 bytes), frameworkOffHeapSize=128.000mb (134217728 bytes),
>>> taskHeapSize=4.125gb (4429184954 bytes), taskOffHeapSize=0 bytes,
>>> networkMemSize=896.000mb (939524110 bytes), managedMemorySize=3.500gb (
>>> 3758096440 bytes), jvmMetaspaceSize=256.000mb (268435456 bytes),
>>> jvmOverheadSize=1024.000mb (1073741824 bytes)}.
>>>
>>
>>>
>>
>>  which are as per defaults.
>>
>> Now, after 25 days we started encountering the following yarn container
>> kill error:
>>
>>> Association with remote system [akka.tcp://flink@...] has failed,
>>> address is now gated for [50] ms. Reason: [Association failed with
>>> [akka.tcp://flink@...]] Caused by: [java.net
>>> <https://urldefense.com/v3/__http://java.net/__;!!EFVe01R3CjU!Ip3Kjw-Zxo6kwXvUVlcRfPhaO5QmiI_MdJ5oBurn-evN9B1gr3XT66Tq95hdv3Dfaw$>
>>> .ConnectException: Connection refused: .../...:37679]
>>> 2020-09-09 00:53:24 INFO Closing TaskExecutor connection
>>> container_e193_1592804717489_149347_01_000011 because: [2020-09-09 00:53
>>> :21.417]Container [pid=44371,containerID=container_e193_1592804717489_149347_01_000011]
>>> is running beyond physical memory limits. Current usage: 12.0 GB of 12
>>> GB physical memory used; 14.4 GB of 25.2 GB virtual memory used. Killing
>>> container.
>>>
>>
>> Yarn container size is 12GB as it is only allowed as a multiple of 3 GB
>> (as per our settings).
>>
>> Now, when the YARN reallocates a new container, the program starts again
>> (without any issues) and after a few hours another container is killed with
>> the same error and the cycle repeats.
>> At this point, I want to debug it as a running process without changing
>> or playing around with various config options for memory as I don't think
>> just to reproduce the error, I want to wait for ~1 month.
>>
>> I have tried to figure out something from Graphite metrics (see
>> attachments):
>> [1]: JVM Heap Memory (First 25 days) -> The memory goes up and after
>> reaching a point goes does and again starts going up. (No container kills
>> were encountered until 09/09/2020, program started on 14/08/2020)
>> [2]: JVM Heap Memory (Recent) -> The memory is still going up but it
>> seems it doesn't even reaches its peak, but instead container is killed
>> before that itself (within a few hours)
>>
>> From [1] and [2], JVM heap memory should not rise up I think, but that
>> doesn't explain container kill in [2] case if JVM heap memory was the issue
>> causing container kill.
>>
>> [3]: Direct Memory and Off heap Memory -> I don't think this is causing
>> the issue as most of the network buffers are free and off heap memory is
>> well below threshold.
>>
>> At this point I thought RocksDB might be the culprit. I am aware that it
>> uses the managed memory limits (I haven't changed any default config) which
>> is completely off heap. But when I see the rocksDB size maintained at
>> location:
>>
>>
>>> /data_4/yarn-nm-local-dir/usercache/root/appcache/application_.../flink-io-a48d1127-58a1-41c5-a5f0-32c5180fe74d/job_0bff1881431b5774c3b496a98febed1a_op_WindowOperator_4061fbe16fb95459a1a8d207644e2e63__4_12__uuid_9fe0b2ff-24bc-4301-8044-3fe8e1b3a3a0/db/
>>
>>
>> It is only 17MB which doesn't seem much. I also took a heap dump
>> of org.apache.flink.yarn.YarnTaskExecutorRunner process but it shows only
>> 30MB of data is being used (not sure what I am missing here as it doesn't
>> match with metrics shown by flink).
>>
>> Although top -p 'pid' (for task manager process) does show RES = 10-12 GB
>> for every container constantly going up and eventually dies.
>>
>> Has someone encountered a similar situation or have guidelines that I can
>> continue with to figure out and debug the issue? Let me know if there is
>> anything else that you might wanna know.
>>
>> --
>> Thanks & Regards
>>
>> Shubham Kumar
>>
>> ------------------------------
>
> The information contained in this e-mail is confidential and/or
> proprietary to Capital One and/or its affiliates and may only be used
> solely in performance of work or services for Capital One. The information
> transmitted herewith is intended only for use by the individual or entity
> to which it is addressed. If the reader of this message is not the intended
> recipient, you are hereby notified that any review, retransmission,
> dissemination, distribution, copying or other use of, or taking of any
> action in reliance upon this information is strictly prohibited. If you
> have received this communication in error, please contact the sender and
> delete the material from your computer.
>
>
>
>
>

-- 
Thanks & Regards

Shubham Kumar