Hi PengHui, More detailed information is needed to decide whether to revert. We have launched it to the production environment, and everything seems fine. If there is any evidence indicating that the issue was introduced by this PR, we can either fix it or revert it. However, we cannot revert it based solely on suspicion.
Thanks, Wenzhi Feng. On 2025/06/18 22:26:26 PengHui Li wrote: > Hi team, > > I am writing to propose an urgent revert of PR #23611 > > - https://github.com/apache/pulsar/pull/23611 > > *Reason for Revert:* > > This PR has introduced a significant performance regression in the Pulsar > broker. > The attached flame graph visually demonstrates increased CPU utilization > and time > spent in the code paths related to DelayedDeliveryTracker and stream > operations. > > While the intention was to optimize memory usage, the current > implementation > appears to have an adverse effect on CPU performance, leading to overall > degraded > broker throughput and increased latency. > > *Impact:* > > This regression is impacting the stability and performance of our Pulsar > clusters, > especially when you have large-scale delayed messages. Reverting the change > will > allow us to restore the previous performance characteristics while we > investigate > a more robust and performant solution for DelayedDeliveryTracker memory > optimization. > > *Proposed Action:* > > I propose to revert PR #23611 as soon as possible to reduce the risk to > other users, > and we can then collectively work on a more thoroughly tested and > performant approach > to optimize the DelayedDeliveryTracker memory. > > Please let me know your thoughts, and if there are any immediate concerns > with this proposed revert. > > Regards, > Penghui >
