[FLINK-27311] Rocksdb mapstate behaves unexpectedly

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Not A Bug
• Affects Version/s: 1.14.4
• Component/s: Runtime / State Backends, Stateful Functions
• Labels: MapState State auto-deprioritized-major rocksdb stale-minor

Environment

Kubernetes

• Flags: Important
• Assignee: Unassigned
• Reporter: Jonathan diamant
• Created: 19/Apr/22 10:38
• Updated: 16/Aug/23 04:10
• Resolved: 16/Aug/23 04:10

Description

We use rocksDb backend for our state and we experience an unexpected behavior. The state we use is MapState<String, List<Object>> and when a restart occurs and the state is being recovered from the last checkpoints, it seems that not all the list corresponding to a certain key has been loaded and only after all the state has been recovered it behaves as expected. 

Our guess is that while recovering the state, rocksdb recovers the state in chunks and loads entries of the map not as a whole even though we expect that for every key, the value (a list) will be loaded as one object at once.