Status

Current state: "Under Discussion"

...

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

Motivation

Today Kafka Streams use a single-thread per task architecture to achieve embarrassing parallelism and good isolation. However there are a couple scenarios where async processing may be preferable:

...

3) Delayed processing. One use case is delaying re-processing (e.g. "delay re-processing this event for 5 minutes") as mentioned in 2), another is for implementing a scheduler: e.g. do some additional operations later based on this processed record. based on Zalando Dublin, for example, are implementing a distributed web crawler. Note that although this feature can be handled in punctuation, it is not well aligned with our current offset committing behavior, which always advance the offset once the record has been done traversing the topology.

Public Interfaces

Briefly list any new interfaces that will be introduced as part of this proposal or any existing interfaces that will be removed or changed. The purpose of this section is to concisely call out the public contract that will come along with this feature.

...

• Binary log format

• The network protocol and api behavior

• Any class in the public packages under clientsConfiguration, especially client configuration

  • org/apache/kafka/common/serialization

  • org/apache/kafka/common

  • org/apache/kafka/common/errors

  • org/apache/kafka/clients/producer

  • org/apache/kafka/clients/consumer (eventually, once stable)

• Monitoring

• Command line tools and arguments

• Anything else that will likely break existing users in some way when they upgrade

Proposed Changes

We should add a mechanism for out-of-order execution with in-order commit - not unlike the dynamic scheduling algorithm/ architecture invented by Tomasulo in the '60s. In that sense, 

...

With this proposal, a message goes through a StreamTask in three stages:

Stage 1: Issue

• wait until there's an empty RS
• Retrieve next record ("instruction") from the source processors, fill in the RS

Stage 2: Execute

• wait until source data is available (async processing completed in prior processor(s))
• start async processing
• mark RS as "ready" on callback (when I/O is finished)

Stage 3: Store & commit

• (this happens in the sink processors) Write result to the output topic,
• Mark input partition/offset "complete"
• If there's a prefix of "completed" or "cancelled" offsets commit the last offset for partition & discard the prefix.

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

Lightweight approach

Just make the commit() mechanism more customizable to users for them to implement multi-threading processing themselves: users can always do async processing in the Processor API by spawning a thread-poll, e.g. but the key is that the offset to be committed should be only advanced with such async processing is done. This is a light-weight approach: we provide all the pieces and tools, and users stack them up to build their own LEGOs.