Hi Robert,

Thanks for your suggestions. We are using the DataStream API and I tried it with disabling it completely, but that didn't help.

I attached the plan and to add some context, it starts with a Kafka source followed by a map operation ( parallelism 4). The next map is the expensive part with a parallelism of 18 which produces a Tuple2 which is used for splitting. Starting here the parallelism is always 2 except the sink with 1. Both resulting streams have two maps, a filter, one more map and are ending with an assignTimestampsAndWatermarks. If there is now a small box in the picture it is a filter operation and otherwise it goes directly to a keyBy, timewindow and apply operation followed by a sink.

If one task manager contains more sub tasks of the expensive map than any other task manager, everything later in the stream is running on the same task manager. If two task manager have the same amount of sub tasks, the following tasks with a parallelism of 2 are distributed over the two task manager.

Interesting is also that the task manager have 6 task slots configured and the expensive part has 6 sub tasks on one task manager but still everything later in the flow is running on this task manager. This also happens if operator chaining is disabled.


On 12.10.2016 17:43, Robert Metzger wrote:
Hi Jürgen,

Are you using the DataStream or the DataSet API?
Maybe the operator chaining is causing too many operations to be "packed" into one task. Check out this documentation page: https://ci.apache.org/projects/flink/flink-docs-master/dev/datastream_api.html#task-chaining-and-resource-groups You could try to disable chaining completely to see if that resolves the issue (you'll probably pay for this by having more serialization overhead and network traffic).

If my suggestions don't help, can you post a screenshot of your job plan (from the web interface) here, so that we see what operations you are performing?


On Wed, Oct 12, 2016 at 12:52 PM, Jürgen Thomann <juergen.thom...@innogames.com <mailto:juergen.thom...@innogames.com>> wrote:


    we currently have an issue with Flink, as it allocates many tasks
    to the same task manager and as a result it overloads it. I
    reduced the amount of task slots per task manager (keeping the CPU
    count) and added some more servers but that did not help to
    distribute the load.

    Is there some way to force Flink to distribute the load/tasks on a
    standalone cluster? I saw that
    <https://issues.apache.org/jira/browse/FLINK-1003> would maybe
    provide what we need, but that is currently not worked on as it seems.


Reply via email to