We currently have a flink 1.8 application deployed on Kinesis Data Analytics using the RocksDB State backend. Our application is joining across 3 different kinesis streams using an interval join. We noticed that our checkpoint sizes continue to increase over time, we eventually have OOM failures writing checkpoints and need to restart the application without restoring from a savepoint.
Does this kind of application require a state TTL on the join operator? I assumed since it was an interval join, events that fell outside of the lower timebound would automatically be expired from the state. Is that a correct assumption? Thanks, Chris