0

I'm using Kafka to distribute load between my AI agents but my servers have different configurations and different rates for processing the input data, a few of them remain mostly idle while others lag behind.

it's because right now Kafka's algorithms assign equal number of partitions to each consumer, for example if my topic has 4 partitions and I have 2 consumers each one subscribes to 2 partitions. I know I can assign consumers to specific partitions but I don't want to limit my configurations to that extent.

is there a way to set a limit for how many partitions a consumer can subscribe to without specifying which partitions exactly?

my consumer is implemented using kafka-python==2.0.2 .

EDIT: as OneCrecketeer kindly mentioned, there is no guarantee that equal partitions will be assigned to each consumer.

aSaffary
  • 793
  • 9
  • 22
  • 1
    Equal partition assignment isn't guaranteed, by the way. One instance could have 4 or 3 partitions assigned as well – OneCricketeer Jul 24 '22 at 16:40
  • I put a java solution below although I realize now you need a python-based solution. Does this code snippet answers your question? https://github.com/confluentinc/confluent-kafka-python/issues/1071 – Svend Jul 25 '22 at 07:10
  • @OneCricketeer thanks, I'll add that to my question as well :) – aSaffary Jul 25 '22 at 08:58

1 Answers1

1

There is no out of the box configuration that allows to assign explicitly a specific number of partitions to each member of the consumer group.

The behavior for assigning partitions to each member of the group is controlled by the partition.assignment.strategy parameter, which at the moment accepts the following values:

  • RangeAssignor (default): a range of consecutive partitions is assigned to each member)
  • RoundRobinAssignor: partitions are assigned one by one to each member, which can result in fairer load distribution when subscribing to several topics
  • StickyAssignor: with best-effort sticky assignments across restart, potentially helping application better re-use local cache
  • CooperativeStickyAssignor variation of the above, reducing the negative impact of a rebalance on throughput.

See this How to assign partitions to your consumers for a high level overview of how this works.

None of those is doing what you're looking for, but you could provide your own implementation of ConsumerPartitionAssignor, taking into account some weight parameter you would have added somehow. See the javadoc linked above for more details.

Note that partition assignment is something that happens on the client side (in one of the consumer currently running), so your implementation must be available in the classpath of the client (not on the broker)

Svend
  • 6,352
  • 1
  • 25
  • 38