This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: [One of "Under Discussion", "Accepted", "Rejected"]

Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]

JIRA: here [Change the link from KAFKA-1 to your own ticket]

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Consumer semantics is very useful for distributed processing of the data, however the granularity of parallelism doesn’t satisfy the scaling need when number of topic partitions < number of consumers. Nowadays Kafka client users would do the capacity planning beforehand to allow 5X ~ 10X future traffic increase. This aims to avoid hitting the future scalability issue at the best effort, but still possible that eventually the traffic goes beyond the original planning, and user has to face the unfortunate online migration. One solution we have considered is to do online partition expanding. The proposal was not continuing to evolve due to its complexity. A second option which also painful is to switch input topic on the fly. As of today, the switch process is manual and cumbersome. 

In the infra cost perspective, pre-define a higher number of partitions will definitely increase the network traffic as more metadata and replication will be needed. Besides extra money paid, the operation overhead increases while maintaining broker cluster in good shape with more topic partitions beyond necessity. It's been a known pain point for Kafka streaming processing scalability which is of great value to be resolved. 

The proposal here, is to decouple the consumption and physical partition count, by making consumers capable of collaborating on the same topic partition. There are a couple of benefits compared with existing model:

1. Data consume and produce scales are no longer coupled. This means we could save money by configuring a reasonable input topic with decent amount of partitions.
2. Better avoid partition level hotkeys. When a specific key is processing really slow, the decoupled key based consumption could bypass it and make progress on other keys.

3. No operation overhead. Users just need to add more consumer/stream capacity to unblock.

For streaming operations such as join, aggregation and so on, the relative order between different keys is not necessary to be maintained, however currently topic partition data must be processed in order for easy offset management. If one hot key is blocking the way, all subsequent data processing will be stopped. 

Public Interfaces

The 

Proposed Changes

Describe the new thing you want to do in appropriate detail. This may be fairly extensive and have large subsections of its own. Or it may be a few sentences. Use judgement based on the scope of the change.

Compatibility, Deprecation, and Migration Plan

• What impact (if any) will there be on existing users?
• If we are changing behavior how will we phase out the older behavior?
• If we need special migration tools, describe them here.
• When will we remove the existing behavior?

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.