Kafka Per-Partition Watermarks

[Andrew Kowpak]

Hi all,

I apologize if this has been discussed to death in the past, but, I'm finding myself very confused, and google is not proving helpful.

Based on the documentation, I understand that if there are idle partitions in a kafka stream, watermarks will not advance for the entire application.  I was hoping that by setting parallelism = the number of partitions that I would be able to work around the issue, but, this didn't work.  I'm totally willing to accept the fact that if I have idle partitions, my windowed partitions won't work, however, I would really like to understand why setting the parallelism didn't work.  If someone can explain, or perhaps point me to documentation or code, it would be very much appreciated.

Thanks.

--

[Elias Levy]

Does your job perform a keyBy or broadcast that would result in data from different partitions being distributed among tasks?  If so, then that would be the cause. 

On Thu, Oct 4, 2018 at 12:58 PM Andrew Kowpak <[hidden email]> wrote:

Hi all,

I apologize if this has been discussed to death in the past, but, I'm finding myself very confused, and google is not proving helpful.

Based on the documentation, I understand that if there are idle partitions in a kafka stream, watermarks will not advance for the entire application.  I was hoping that by setting parallelism = the number of partitions that I would be able to work around the issue, but, this didn't work.  I'm totally willing to accept the fact that if I have idle partitions, my windowed partitions won't work, however, I would really like to understand why setting the parallelism didn't work.  If someone can explain, or perhaps point me to documentation or code, it would be very much appreciated.

Thanks.

--

[Taher Koitawala]

Hey Andrew,

          We face the same problem in our application where some of the kafka partitions are empty. In this case what we do is use the rebalance() method on the source streams.

Ex: DataStream<T> srcStream= Env.addSource(new FlinkKafkaConsumer09<>(topic,Ser,props));

SrcStream.rebalance();

Since rebalance sends records all tasks in a round robin fashion we do not have to worry about empty partitions and also about mapping our parallelism to kafka partitions. After rebalance you can have your parallelism higher than the number of kafka partition and windowing will work just fine.

Thanks, 

Taher Koitawala 

On Fri 5 Oct, 2018, 1:28 AM Andrew Kowpak, <[hidden email]> wrote:

Hi all,

I apologize if this has been discussed to death in the past, but, I'm finding myself very confused, and google is not proving helpful.

Based on the documentation, I understand that if there are idle partitions in a kafka stream, watermarks will not advance for the entire application.  I was hoping that by setting parallelism = the number of partitions that I would be able to work around the issue, but, this didn't work.  I'm totally willing to accept the fact that if I have idle partitions, my windowed partitions won't work, however, I would really like to understand why setting the parallelism didn't work.  If someone can explain, or perhaps point me to documentation or code, it would be very much appreciated.

Thanks.

--

[Andrew Kowpak]

Yes, my job does do a keyBy.  It never occurred to me that keyBy would distributed data from different partitions to different tasks, but, now that you mention it, it actually makes perfect sense.  Thanks you for the help.

On Thu, Oct 4, 2018 at 5:11 PM Elias Levy <[hidden email]> wrote:

Does your job perform a keyBy or broadcast that would result in data from different partitions being distributed among tasks?  If so, then that would be the cause. 

On Thu, Oct 4, 2018 at 12:58 PM Andrew Kowpak <[hidden email]> wrote:

Hi all,

I apologize if this has been discussed to death in the past, but, I'm finding myself very confused, and google is not proving helpful.

Based on the documentation, I understand that if there are idle partitions in a kafka stream, watermarks will not advance for the entire application.  I was hoping that by setting parallelism = the number of partitions that I would be able to work around the issue, but, this didn't work.  I'm totally willing to accept the fact that if I have idle partitions, my windowed partitions won't work, however, I would really like to understand why setting the parallelism didn't work.  If someone can explain, or perhaps point me to documentation or code, it would be very much appreciated.

Thanks.

--

--