Spread Kafka sink tasks over different nodes

[Dongwon Kim-2]

Hi,

We have a standalone cluster where 1 JM and 7 TMs are running on 8 servers.

We have total 224 cores (as each TM has 32 slots) and we want to use all slots for a single streaming job.

The single job roughly consists of the following three types of tasks:

- Kafka source tasks (Parallelism : 7 as the number of partitions in the input topic is 7)

- Session window tasks (Parallelism : 224)

- Kafka sink tasks (Parallelism : 7  as the number of partitions in the output topic is 7)

We want 7 sources and 7 sinks to be evenly scheduled over different nodes.

Source tasks are scheduled as wanted (see "1 source.png").

However, sink tasks are scheduled on a single node (see "2 sink.png").

As we use the whole standalone only for a single job, this scheduling behavior causes the output of all the 224 session window tasks to be sent to a single physical machine.

Is it because locality is only considered in Kafka source?

I also check that different partitions are taken care by different brokers for both of the input topic and the output topic in Kafka.

Do I miss something in order to spread Kafka sink tasks over different nodes?

Best,

- Dongwon

[Aljoscha Krettek]

Hi Dongwon,

I think there is currently no way of ensuring that tasks are spread out across different machines because the scheduling logic does not take into account what machine a slot is on. I currently see two workarounds:

 - Let all operations have the same parallelism and only have 8 slots in your cluster in total
 - Let your sinks have parallelism 224 (same as the windows). I think multiple sinks writing to the same Kafka partition should not be a problem. Unless that's a problem in your setup, of course.

What do you think?

Best,
Aljoscha

> On 2. Mar 2018, at 06:55, Dongwon Kim <[hidden email]> wrote:
>
> Hi,
>
> We have a standalone cluster where 1 JM and 7 TMs are running on 8 servers.
> We have total 224 cores (as each TM has 32 slots) and we want to use all slots for a single streaming job.
>
> The single job roughly consists of the following three types of tasks:
> - Kafka source tasks (Parallelism : 7 as the number of partitions in the input topic is 7)
> - Session window tasks (Parallelism : 224)
> - Kafka sink tasks (Parallelism : 7  as the number of partitions in the output topic is 7)
>
> We want 7 sources and 7 sinks to be evenly scheduled over different nodes.
> Source tasks are scheduled as wanted (see "1 source.png").
> <1 source.png>
>
> However, sink tasks are scheduled on a single node (see "2 sink.png").
> <2 sink.png>
>
> As we use the whole standalone only for a single job, this scheduling behavior causes the output of all the 224 session window tasks to be sent to a single physical machine.
>
> Is it because locality is only considered in Kafka source?
> I also check that different partitions are taken care by different brokers for both of the input topic and the output topic in Kafka.
>
> Do I miss something in order to spread Kafka sink tasks over different nodes?
>
> Best,
>
> - Dongwon

[Dongwon Kim-2]

Hi Aljoscha and Robert,

You guys are right.
I resubmit the application with # session window tasks equal to # Kafka sink tasks.
I never thought that multiple different Kafka tasks can write to the same partition.

Initially, I do not set the default parallelism and I explicitly set # partitions of each stage.
This decision naturally led me to set # Kafka sink stage to # of Kafka partitions.

Sorry for the confusion :-)

Best,

Dongwon
       

> 2018. 3. 6. 오후 10:28, Aljoscha Krettek <[hidden email]> 작성:
>
> Hi Dongwon,
>
> I think there is currently no way of ensuring that tasks are spread out across different machines because the scheduling logic does not take into account what machine a slot is on. I currently see two workarounds:
>
> - Let all operations have the same parallelism and only have 8 slots in your cluster in total
> - Let your sinks have parallelism 224 (same as the windows). I think multiple sinks writing to the same Kafka partition should not be a problem. Unless that's a problem in your setup, of course.
>
> What do you think?
>
> Best,
> Aljoscha
>
>> On 2. Mar 2018, at 06:55, Dongwon Kim <[hidden email]> wrote:
>>
>> Hi,
>>
>> We have a standalone cluster where 1 JM and 7 TMs are running on 8 servers.
>> We have total 224 cores (as each TM has 32 slots) and we want to use all slots for a single streaming job.
>>
>> The single job roughly consists of the following three types of tasks:
>> - Kafka source tasks (Parallelism : 7 as the number of partitions in the input topic is 7)
>> - Session window tasks (Parallelism : 224)
>> - Kafka sink tasks (Parallelism : 7  as the number of partitions in the output topic is 7)
>>
>> We want 7 sources and 7 sinks to be evenly scheduled over different nodes.
>> Source tasks are scheduled as wanted (see "1 source.png").
>> <1 source.png>
>>
>> However, sink tasks are scheduled on a single node (see "2 sink.png").
>> <2 sink.png>
>>
>> As we use the whole standalone only for a single job, this scheduling behavior causes the output of all the 224 session window tasks to be sent to a single physical machine.
>>
>> Is it because locality is only considered in Kafka source?
>> I also check that different partitions are taken care by different brokers for both of the input topic and the output topic in Kafka.
>>
>> Do I miss something in order to spread Kafka sink tasks over different nodes?
>>
>> Best,
>>
>> - Dongwon
>