Status

Current state: Under Discussion

Discussion thread: here

JIRA: here

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

Motivation

Now TCP_NODELAY socket option is always enabled and large number of topic-partitions on one broker causes burst host's packets/sec metric. This can cause service degradation in cases where you are using a cloud traffic shaper or other network controls.

For example, for test cluster with 4 brokers and 30 000 topic-partitions: 

~99.999% of all packets is inter broker messages. Size of packets 40-160 bytes.

In our production cluster ~27k topic-partitions at 16 brokers generate ~500 000 tcp packets per sec (IN and OUT summary).

More how reproduce, research and fix this issue in description of JIRA ticket.

Public Interfaces

New broker property (boolean, default value = true):

socket.tcp.no.delay

Proposed Changes

New boolean property in Kafka Config that forwarded into Acceptor class in SocketServer.scala. Pull request small enough: just 12 lines of code.

Compatibility, Deprecation, and Migration Plan

New changes has 100% compatibility with all kafka releases and may be backported to any previous release with no changes. No deprecation or migration plan are needed. The default value is the current behaviour.

Rejected Alternatives

At this moment all new connection in broker are created with enabled TCP_NODELAY socket option. Because "true" value is hardcoded since 2011 year. There is no way to change this option outside.