Github user EronWright commented on the issue:
https://github.com/apache/flink/pull/5427
I feel that we're not addressing the core issue that we're trying to fix.
1. A new job starts up with checkpointing enabled and a hook-based source.
2. The source begins to consume events, causing some external state to
become mutated.
3. _Before the first checkpoint_, a task throws an exception, causing a
global restart.
4. Since the hook has no opportunity to rewind the external state to
initial conditions, data loss occurs.
The above is a special case. In the normal case, one or more checkpoints
have occurred before the restart occurs, and so the hook's `restore` method is
effective.
---
