We are using the rocksDB state backend. We had not activated incremental checkpointing, but in the course of debugging this, we ended up doing so, and also moving back to S3 from EFS as it appeared that EFS was introducing large latencies. I will attempt to provide some profiler data as we are able to analyze further.

--
Jared Stehler
Chief Architect - Intellify Learning
o: 617.701.6330 x703



On Jul 3, 2017, at 6:02 AM, Stefan Richter <s.richter@data-artisans.com> wrote:

Hi,

I have two quick questions about this problem report:

1) Which state backend are you using?
2) In case you are using RocksDB, did you also activate incremental checkpointing when moving to Flink 1.3.

Another thing that could be really helpful, if possible, can you attach a profiler/sampling to your job manager and figure out the hotspot methods where most time is spend? This would be very helpful as a starting point where the problem is potentially caused.

Best,
Stefan

Am 29.06.2017 um 18:02 schrieb Jared Stehler <jared.stehler@intellifylearning.com>:

We’re seeing our external checkpoints directory grow in an unbounded fashion… after upgrading to Flink 1.3.  We are using Flink-Mesos.

In 1.2 (HA standalone mode), we saw (correctly) that only the latest external checkpoint was being retained (i.e., respecting state.checkpoints.num-retained default of 1)

The Mesos-agent running the Job Manager ends up with a really high load and ends up getting unresponsive….  Interestingly enough, there is not much CPU or Memory pressure… so it is suggesting to us that its IO or Network bound.  But nothing jumps out at us (using iostat/netstat).  The biggest difference seems to be external checkpoints not getting cleaned up/discarded.  What might cause that?

ubuntu@ip-10-80-52-176:/mnt/shared/flink/ext-checkpoints$ top
top - 13:47:41 up 16:31,  1 user,  load average: 25.85, 25.62, 25.43
Tasks: 297 total,   1 running, 296 sleeping,   0 stopped,   0 zombie
%Cpu(s):  0.3 us,  0.0 sy,  0.0 ni, 98.8 id,  0.7 wa,  0.0 hi,  0.0 si,  0.0 st
KiB Mem:  32948204 total, 23974844 used,  8973360 free,   144572 buffers
KiB Swap:        0 total,        0 used,        0 free.  7752480 cached Mem

We specify Mesos agent attributes to ensure that our Flink containers are allocated to only a subset of the Mesos slaves…   However, we do end up with the Flink JobManager container running on the same physical instance as multiple task manager containers. We are running 65 task managers with 2 slots each, and ~70 jobs currently on the cluster.

We use AWS EFS (https://aws.amazon.com/efs/) mounted on all Mesos boxes to store recovery, checkpoint, external checkpoint and save point directories.


        executionEnvironment.enableCheckpointing(TimeUnit.SECONDS.toMillis(30));

        CheckpointConfig config = executionEnvironment.getCheckpointConfig();
        config.enableExternalizedCheckpoints(ExternalizedCheckpointCleanup.RETAIN_ON_CANCELLATION);
        config.setMinPauseBetweenCheckpoints(TimeUnit.SECONDS.toMillis(5));

        executionEnvironment.getConfig().setGlobalJobParameters(params);
        executionEnvironment.getConfig().setAutoWatermarkInterval(watermarkInterval.getValue());
        executionEnvironment.getConfig().setCodeAnalysisMode(CodeAnalysisMode.HINT);

        executionEnvironment.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);

        // fail the job if it restarts more than 3 times in 1 minute, with 10 second delay
        executionEnvironment.setRestartStrategy(RestartStrategies.failureRateRestart(3,
                Time.minutes(2), Time.seconds(1)));

        executionEnvironment.getConfig().setLatencyTrackingInterval(30000);


Would appreciate any insights you might have on this. 

Thanks

--
Jared Stehler
Chief Architect - Intellify Learning
o: 617.701.6330 x703