Hi all, A few clarifications on the patch:
1. #1716 doesn't introduce a new config option. Making the limit independently configurable was discussed, and ztzg explained why it isn't needed. it simply tracks jute.maxbuffer: https://github.com/apache/zookeeper/pull/1716#issuecomment-866691807 So there is nothing to deprecate later: when CloseSessionTxn goes away, the check goes away with it. The point of the patch is crash protection, not a new mechanism. 2. What it does introduce is a new error code, which old clients won't recognize. That's a real trade-off, but the alternative is the whole cluster going down. An unknown error on the one misbehaving client is harmless in comparison. 3. On the -1: @eolivelli seems to have stepped back from it in Aug 2024. "I think that we can resume this patch ... I don't have a better proposal at the moment": https://github.com/apache/zookeeper/pull/1716#issuecomment-2296414881 On process: the 2024 revival already turned into the non-fuzzy snapshot discussion and then went quiet for lack of participants, while the production problem stayed. I'd suggest keeping this thread to the blockers for merging the fix - a broader discussion will most likely end the same way. On ZOOKEEPER-4874 being small: technically perhaps, but conceptually it reworks durability, and I'd wait a few months after such a release before running it in production anyway, so it can't be the near-term answer. Andor, as for our numbers and use case: I'd rather not build the case on our specifics. Sharing details would have to go through internal approval, and that takes time. What I can say: I run ZooKeeper for many products in a large company, two independent products hit this issue, and there are other reports of it too. We used to run an old internally patched version and recently moved to the latest open-source release, wanting to stay with the community, and this is the first thing we hit. Let's focus on stopping the production crashes. Dmitrii > Hi all, > > We've hit ZOOKEEPER-4306 (CloseSessionTxn exceeding jute.maxbuffer takes down > the whole ensemble) twice in production, in two independent deployments — > I've also left a comment on the JIRA. > > PR #1716 has been open since 2021 and was rebased in Oct 2024 with a > branch-3.9 backport (#2201), but both seem to have stalled since. > > Is anything still blocking them? If the 2021 review comments are the > remaining concern, we'd be glad to help address them > > https://issues.apache.org/jira/browse/ZOOKEEPER-4306 > https://github.com/apache/zookeeper/pull/1716 > https://github.com/apache/zookeeper/pull/2201
