Resurfacing this as it seems (to me) an important correction to be made. Would love your opinions on this.
Thanks! Asaf On Wed, May 11, 2022 at 1:16 PM Asaf Mesika <asaf.mes...@gmail.com> wrote: > Hello all, > > I have a question regarding a feature introduced in 2.6.0 called Negative > Acknowledgement in Batch Level index, described in PIP-54 > <https://github.com/apache/pulsar/wiki/PIP-54:-Support-acknowledgment-at-batch-index-level> > . > > In short: Before this feature, if you would write in batches (in producer) > to Pulsar (which makes sense most would for performance reasons), then a > consumer sending a negative acknowledgment for 1 message in the batch (say > we have 500 in a single batch), would cause the consumer to get the all > messages in the batch (500) redelivered as if they were not acknowledge. > > PIP-54 fixes that by keeping the ack per message in a batch using a bitmap > index. The caveats are more memory consumption since the broker keeps those > bitmaps in-memory for any inflight batch. > > With PIP-54 Pulsar IMO becomes "complete" in that it acts the way you > would expect it to be, in normal circumstances (as opposed to disconnects, > machine crashing, etc). > > This feature was introduced in 2.6.0, roughly 2 years ago, and is off by > default. > > I was wondering what the developers community thoughts on turning it on by > default? Has anyone experienced any performance degradation to it? Have you > turned it on in your clusters? > > Thanks a lot for your time. > > Asaf Mesika >