>> This version worked a lot better for us, but as long as we're unable to reproduce and fix the missing messages, we're staining on 2.20 for now.<<<
I have been working closely with some customers running huge workloads, and I have never seen any message loss.. I believe you probably ran on a split brain situation (If you use replication), to me a message loss means you sent a message, restarted the broker and the message never recovered.. I have never seen that. A split brain should be fixed by keeping 3 pairs of replicas, or use zooKeeper. On Wed, Oct 26, 2022 at 5:02 PM Walter de Boer <walterdeb...@dbso.nl> wrote: > > @Clebert, thanks for the update! > > @Caspar, with 2.20 we do see memory issues that I believe are solved in > later versions. But we didn't lose messages the same way we did with > 2.24. With 2.20 we're monitoring for AMQ222162 after a crash. We then > recover the old replica's just in case replication failed for some time > because of the memory issues. We have our own duplicate message > detection in our clients, so getting them twice is no prolem for us. But > normally replicaiton works as expected and I guess recovering is not > even needed. > > We saw crashes with 2.20 on G1GC and this week also on ZGC, but I know > there were some memory related fixes in later releases. We now prevent > crashes by restarting every few day's, but I would like to go up again > to 2.26. This version worked a lot better for us, but as long as we're > unable to reproduce and fix the missing messages, we're staining on 2.20 > for now. > > -- Clebert Suconic