[FLINK-6533] Duplicated registration of new shared state when checkpoint confirmations are still pending

• Type: Sub-task
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 1.3.0
• Fix Version/s: 1.3.0
• Component/s: Runtime / State Backends
• Assignee: Stefan Richter
• Reporter: Stefan Richter
• Created: 11/May/17 08:39
• Updated: 14/May/17 12:25
• Resolved: 14/May/17 12:25

Description

Each incremental RocksDB checkpoint n is registering new and existing shared state with the SharedStateRegistry when it completes. Only then, the backend is notified and all following checkpoints (n+x) can reference the new state in checkpoint n.

However, when a checkpoint n+1 is already starting before n was confirmed to the backend, n+1 can assume some files as new, which were already contained in n. It will upload the file to DFS again, creating a new state handle.

Then, once n+1 completes, it could to register some state as new, which was previously registered already by n, without n+1 knowing of this. Currently this violates a precondition check, that the reference count for state that is assumed as new is 1.

While we cannot prevent duplicate uploads, we must resolve this situation in the SharedStateREgistry