That would reduce the size of the checkpoints but as far as I know not the savepoints. If i understand correctly that is still a copy of the entire system state. So deploying a new version of the application will be a daunting proposition involving saving many gigabytes of data to external storage. and restarts that will take a very long time. Also our flink workers are not really scoped to these kind of storage requirements
> 2 Have you considered using queryable state vs. also keeping the list of events in Cassandra?
We looked at it before and at the time it was still somewhat experimental and somewhat immature with regards to handling failure scenarios. And it would require all the state to reside in flink. Which would again lead to long restarts when creating savepoints
> 3. Depending on what you need the list of events for, often you can apply a streaming algorithm to get good-enough (approximate) results without storing complete state.
Hyperloglog etc makes a lot of sense in many scenario's, but unfortunately this is not one of them. :-(