You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 14 Next »

Status

Current stateUnder Discussion

Discussion thread: [DISCUSS] KIP-281: ConsumerPerformance: Increase Polling Loop Timeout and Make It Reachable by the End User

Vote thread: [VOTE] KIP-281: ConsumerPerformance: Increase Polling Loop Timeout and Make It Reachable by the End User

JIRA Unable to render Jira issues macro, execution error.

PULL REQUEST: https://github.com/apache/kafka/pull/4818

Motivation

ConsumerPerformance fails to consume all messages on topics with large number of partitions due to a relatively short default polling loop timeout (1000 ms) that is not reachable and modifiable by the end user. 

Demo: Create a topic of 10 000 partitions, send a 50 000 000 of 100 byte records using kafka-producer-perf-test and consume them using kafka-consumer-perf-test (ConsumerPerformance). You will likely notice that the number of records returned by the kafka-consumer-perf-test is many times less than expected 50 000 000.

As the result, in some rough cases, it may take a long enough time to process/iterate through the records polled in batches, thus the time may exceed the default hardcoded polling loop timeout (1 second), forcing the utility to stop it's execution much earlier than it should to.

This leads to confusion, poor and unexpected test results and is probably not what we want. Thus:

  • Increasing polling loop timeout is probably a good thing to handle this case
  • Increasing it too much by default (say, 5 minutes) may lead to a slower feedback from the tool in some cases 
  • Making polling loop timeout optionally changeable is a good thing since it would give us more control for the corner cases
  • Printing a warning message in case the timeout is exceeded should also be a good thing to know what's going on  

Public Interfaces

Providing the command line tool kafka-consumer-perf-test.sh with an optional --polling-loop-timeout parameter: 

parameter namedefaults totypetime unitusage
--timeout10000Longmsoptional

Proposed Changes

  • Increasing polling loop timeout in ConsumerPerformance implementation up to 60 000 ms 
  • Making polling loop timeout reachable and modifiable by the end user from the command line kafka-consumer-perf-test tool
  • If the polling loop timeout is exceeded, print a warning message

Proposed changes in code available in this PR: https://github.com/apache/kafka/pull/4818

Compatibility, Deprecation, and Migration Plan

  • Positive impact on existing users since the utility generates more accurate results 
  • The proposed parameter is optional, will not break anything

Rejected Alternatives

None.

 

  • No labels