Hi folks! I noticed that my flinkapp (deployed on AWS Managed Apache Flink) performs always a checkpoint with data size of 2.60MB. Incremental checkpoint is set, in fact Full Checkpoint size is bigger. The lastCheckpointSize doesn't decrease even if there isn't flow of messages into the app and there arent' timers on. What could be checkedpointed that changes each time? I have both ValueState with large, complex objects and ValueState with simple booleans. I have several ValueState in a single KeyedProcessFunction. An input record for the KeyedProcessFunction can cause the filling of only one of the ValueState

the only way this would decrease is when state compaction happens from RocksDB. BTW, 2.6MB is not much. Is there a concern about this? How long are your checkpoints taking? Keep in mind your sources and sinks also generally have state about them even if you're not explicitly accessing them

2.60 MB should be the difference between a checkpoint data and the follow checkpoint data, because incremental checkpoint is set. So I don't understand the reason why you speak about state compaction

i'm not sure, perhaps the KeyedProcessFunction has some metadata state that it needs to maintain. I don't recall ever seeing a checkpoint of 0.0 MB. I'll restate the question, is there a reason for concern here?

I am investigating an ever-increasing checkpoint size, which is why I had asked myself this question

Are you manually clearing your state in your process function?

What are your settings for

rocksdb.write-buffer-size

also is compaction filtering enabled?

I'm clearing state manually and also using cleaning during RocksDB compaction filter process (so setting a TTL)