Status

Current state: "Under DiscussionAccepted"

Discussion thread: here  , and here

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).

...

Even though client will receive the new leader information in the ProduceResponse & FetchResponse when leader changes, but same as the existing behaviour of the Kafka Java client, it will request expedited metadata-refresh done asynchronously. Since leadership change will likely affect many partitions, so future requests to such partitions will benefit from thisthe upto date leadership information, and reduce requests going to old leaders.

For Produce, if new leader info is available in the response along with errors(NOT_LEADER_OR_FOLLOWER or FENCED_LEADER_EPOCH) that advances client's view of the leadership, client would no longer back off up to RETRY_BACKOFF_MS_CONFIG before retrying the failed batch. This immediate retry is appealing as the client is going to retry on a different broker and it is likely to succeed because it is retrying on a newer leader. On the other hand, subsequent retries to the same new LeaderEpoch should still continue to be subject to clients' backoff strategy.

...

Performance was tested using the kafka-producer-perf-test.sh script and reassigning leadership of all partitions of a 100 partition topic. We see an end-to-end reduction in the p99.9 produce latency of the overall run of 88%, from 1675ms to 215ms (average of 3 runs). We hypothesize the residual latency is due to metadata convergence on the servers, this is evident in the results for the rejected alternative, which performs a full metadata refresh but eliminates the retry backoff to the new leader. This experiment showed an average latency of 3022ms which is higher than the baseline, we hypothesize this is due to the high variance server side convergence introduces to metadata latencyDigging deeper into the baseline, & rejected alternate, metadata RPC becomes slow due to slower produce RPCs ahead of it to a given broker. This results into metadata refresh being slow for baseline & rejected alternate over all, resulting into higher latencies. KIP-951 doesn't rely on the Metadata RPC for the new leader information, hence sees better latencies.

Baseline

Run 1: 40000000 records sent, 99997.750051 records/sec (95.37 MB/sec), 12.56 ms avg latency, 8087.00 ms max latency, 6 ms 50th, 8 ms 95th, 12 ms 99th, 2967 ms 99.9th.
Run 2: 40000000 records sent, 99998.250031 records/sec (95.37 MB/sec), 15.51 ms avg latency, 11652.00 ms max latency, 6 ms 50th, 8 ms 95th, 13 ms 99th, 859 ms 99.9th.
Run 3: 40000000 records sent, 99998.000040 records/sec (95.37 MB/sec), 8.63 ms avg latency, 3224.00 ms max latency, 6 ms 50th, 8 ms 95th, 14 ms 99th, 1201 ms 99.9th.

...