Status

Current state: ["Under Discussion"]

Discussion thread: here

JIRA: KAFKA-15777

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

Motivation

When remote storage is enabled on a topic, and the consumer is doing backfill, then the data is read from the remote storage. Currently, Kafka remote storage supports only reading the first remote partition in the given FETCH request. The default value of max.partition.fetch.bytes  is configured to 1 MB. Reading one MB of data per FETCH request, increases the round-trip time for the consumer and the cloud storages are often tuned to read data in chunks of 4 MB. 

The proposal is to introduce remote.max.partition.fetch.bytes  in the Consumer config which allows the user to tune the value depending on their storage plugin. The user might want to optimize the number of calls to remote storage vs the amount of bytes returned back to the client in the FETCH response.

Public Interfaces

• Introduce new remote.max.partition.fetch.bytes consumer config, with default value to 1 MB.
• Bump FetchRequest to v18 to add the RemotePartitionMaxBytes field to propagate the value configured on the consumer to the broker. 

• 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

The proposal is to introduce a new Consumer config: remote.max.partition.fetch.bytes to configure the number of bytes returned from remote storage in a FETCH request. The reason for not re-using the existing the max.partition.fetch.bytes is that:

1. The default value of fetch.max.bytes  is set to 50 MB and max.partition.fetch.bytes is 1 MB.
2. If the user increases the max.partition.fetch.bytes value to 4 MB, then it applies for all the partitions in the FETCH requests. (ie) Reading the data from local storage too.
3. Assume that the consumer is reading data from a topic with 64 partitions and 16 partition leaders are co-located in a single broker. 
   1. Broker allocates 12 instances of 4 MB buffers (12 x 4 = 50 MB) which might trigger the Young Gen and Old Gen GC when there are many FETCH requests from different consumers.
   2. Broker does not use BufferPool to allocate the buffers. 
   3. Allocating big byte-buffers have direct impact on GCs.
4. Remote read gets triggered for only one partition in a given FETCH request. And, we want to allocate the big byte-buffer (4 MB) only for remote read requests.

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?

Test Plan

Describe in few sentences how the KIP will be tested. We are mostly interested in system tests (since unit-tests are specific to implementation details). How will we know that the implementation works as expected? How will we know nothing broke?

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.