Hi, Jai. Could you share your configuration about the checkpoint (interval, min-pause, and so on) and the checkpoint details in the Flink UI ? I guess the delay of the checkpoint may be related to the last checkpoint completion time as you could see in the CheckpointRequestDecider#chooseRequestToExecute. Maybe your checkpoint will last longer every 3rd or 4th checkpoints due to the flush mechanism of rocksdb?
Best, Hangxiang. On Wed, Jun 15, 2022 at 6:27 AM Jai Patel <jai.pa...@cloudkitchens.com> wrote: > We've noticed a spike in the start delays in our incremental checkpoints > every 15 minutes. The Flink job seems to start out smooth, with > checkpoints in in the 15s range and negligible start delays. Then every > 3rd or 4th checkpoint has a long start delay (~2-3 minutes). Teh > checkpoints in between have negligible start delays and are fast. So: > > 2-3 fast with negligible start delay, total time 15-30s > 1-2 slow with 2-3 minute start delay, total time 15-30s longer than the > start delay. > > What could cause this? We have a couple output topics that are > EXACTLY_ONCE, but I switched them to AT_LEAST_ONCE and continued to see the > behavior. > > Thanks. > Jai >