GlobalIndexAssigner crashes on a late bootstrap key after an unaligned checkpoint

[LuciferYang]

Search before asking

• I searched in the issues and found no similar issues.

Paimon version

master (1.5-SNAPSHOT)

Compute Engine

Flink (cross-partition upsert / global dynamic-bucket sink) with unaligned checkpoints.

Minimal reproduce step

1. Run a cross-partition upsert sink (global index assigner) with execution.checkpointing.unaligned.enabled = true.
2. Have a checkpoint land during the bootstrap-drain window, so a KEY_PART record is still queued when the barrier triggers.

What doesn't meet your expectations?

The assigner crashes. GlobalIndexAssigner.bootstrapKey has checkArgument(inBoostrap()), which throws (and bootstrapKeys is already null, so it would NPE) when a KEY_PART record arrives after the bootstrap phase has ended. Under unaligned checkpoints the barrier overtakes buffered records, so prepareSnapshotPreBarrier runs endBoostrap (setting bootstrap=false, nulling bootstrapKeys) while KEY_PART records are still queued; those late records then hit bootstrapKey with bootstrap == false and fail the job, which can re-fail on restart in the same window. Aligned checkpoints never hit this, because barrier alignment processes every pre-barrier KEY_PART before endBoostrap.

Anything else?

No response

Are you willing to submit a PR?

• I'm willing to submit a PR!