[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-729870471 Great work @chia7712 ! With this and #9547, we have converted all of the protocols, which was a huge community effort!

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-729258346 Here are a couple additional test runs. This was on Ubuntu 20 (ami-00831fc7c1e3ddc60). The machine type was m5a.xlarge with 200GB gp2 EBS storage. One instance was running the

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-729081330 @chia7712 Thanks for the updates, looks promising. I'm more inclined to trust your results on linux. I have been testing on macos.Let me see if I can reproduce in a better

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-728669194 @chia7712 It might be worth checking the fancy new `toSend` implementation. I did a quick test and found that gc overhead actually increased with this change even though the

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-728643380 I think the large difference in latency in my test is due to the producer's buffer pool getting exhausted. I was looking at the "bufferpool-wait-ratio" metric exposed in the

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-728545356 It would be helpful if someone can reproduce the tests I did to make sure it is not something funky in my environment.

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-728486374 Yeah, there is something strange going on, especially in regard to latency. Running the same producer performance test, I saw the following: ``` Patch: 25000

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-728383115 Posting allocation flame graphs from the producer before and after this patch: ![Screen Shot 2020-11-16 at 3 02 42

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-724862848 For what it's worth, I think we'll get back whatever we lose here by taking `Struct` out of the serialization path.

[GitHub] [kafka] hachikuji commented on pull request #9401: KAFKA-9628 Replace Produce request/response with automated protocol

hachikuji commented on pull request #9401: URL: https://github.com/apache/kafka/pull/9401#issuecomment-718828042 @chia7712 One thing that would be useful is running the producer-performance test, just to make sure the the performance is inline. Might be worth checking flame graphs as