[
https://issues.apache.org/jira/browse/KAFKA-6037?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Guozhang Wang updated KAFKA-6037:
---------------------------------
Description:
Today the downstream sub-topology's parallelism (aka the number of tasks) are
purely dependent on the upstream sub-topology's parallelism, which ultimately
depends on the source topic's num. partitions. However this does not work
perfectly with dynamic scaling scenarios.
Imagine if your have a simple aggregation application, it would have two
sub-topologies cut by the repartition topic, the first sub-topology would be
very light as it reads from input topics and write to repartition topic based
on the agg-key; the second sub-topology would do the actual work with the agg
state store, etc, hence is heavy computational. Right now the first and second
topology will always have the same number of tasks as the repartition topic
num.partitions is defined to be the same as the source topic num.partitions, so
to scale up we have to increase the number of input topic partitions.
One way to improve on that, is to use a default large number for repartition
topics and also allow users to override it (either through DSL code, or through
config). Doing this different sub-topologies would have different number of
tasks, i.e. parallelism units. In addition, users may also use this config to
"hint" the DSL translator to NOT create the repartition topics (i.e. to not
break up the sub-topologies) when she has the knowledge of the data format.
> Make Sub-topology Parallelism Tunable
> -------------------------------------
>
> Key: KAFKA-6037
> URL: https://issues.apache.org/jira/browse/KAFKA-6037
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Reporter: Guozhang Wang
> Labels: needs-kip
>
> Today the downstream sub-topology's parallelism (aka the number of tasks) are
> purely dependent on the upstream sub-topology's parallelism, which ultimately
> depends on the source topic's num. partitions. However this does not work
> perfectly with dynamic scaling scenarios.
> Imagine if your have a simple aggregation application, it would have two
> sub-topologies cut by the repartition topic, the first sub-topology would be
> very light as it reads from input topics and write to repartition topic based
> on the agg-key; the second sub-topology would do the actual work with the agg
> state store, etc, hence is heavy computational. Right now the first and
> second topology will always have the same number of tasks as the repartition
> topic num.partitions is defined to be the same as the source topic
> num.partitions, so to scale up we have to increase the number of input topic
> partitions.
> One way to improve on that, is to use a default large number for repartition
> topics and also allow users to override it (either through DSL code, or
> through config). Doing this different sub-topologies would have different
> number of tasks, i.e. parallelism units. In addition, users may also use this
> config to "hint" the DSL translator to NOT create the repartition topics
> (i.e. to not break up the sub-topologies) when she has the knowledge of the
> data format.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
