Hi Shachar Why do we see data that is older from lateness configuration There might existed three reasons:
1. RocksDB really still need that file in current checkpoint. If we upload one file named as 42.sst at 2/4 at some old checkpoint, current checkpoint could still include that 42.sst file again if that file is never be compacted since then. This is possible in theory. 2. Your checkpoint size is large and checkpoint coordinator could not remove as fast as possible before exit. 3. That file is created by a crash task manager and not known to checkpoint coordinator. How do I know that the files belong to a valid checkpoint and not a checkpoint of a crushed job - so we can delete those files You have to call Checkpoints#loadCheckpointMetadata[1] to load latest _metadata in checkpoint directory and compare the file paths with current files in checkpoint directory. The ones are not in the checkpoint meta and older than latest checkpoint could be removed. You could follow this to debug or maybe I could write a tool to help know what files could be deleted later. [1] https://github.com/apache/flink/blob/693cb6adc42d75d1db720b45013430a4c6817d4a/flink-runtime/src/main/java/org/apache/flink/runtime/checkpoint/Checkpoints.java#L96 Best Yun Tang ________________________________ From: Shachar Carmeli <carmeli....@gmail.com> Sent: Tuesday, April 7, 2020 16:19 To: user@flink.apache.org <user@flink.apache.org> Subject: Flink incremental checkpointing - how long does data is kept in the share folder We are using Flink 1.6.3 and keeping the checkpoint in CEPH ,retaining only one checkpoint at a time , using incremental and using rocksdb. We run windows with lateness of 3 days , which means that we expect that no data in the checkpoint share folder will be kept after 3-4 days ,Still We see that there is data from more than that e.g. If today is 7/4 there are some files from the 2/4 Sometime we see checkpoints that we assume (due to the fact that its index number is not in synch) that it belongs to a job that crushed and the checkpoint was not used to restore the job My questions are Why do we see data that is older from lateness configuration How do I know that the files belong to a valid checkpoint and not a checkpoint of a crushed job - so we can delete those files
