It seems that
FLINK-10674 has not been ported to the Blink planner.
Because state clean up happens in processing time, it might be the case that retractions are arriving after the state has been cleaned up. Before these changes, a new accumulator was created and invalid retraction messages were emitted. This change drops retraction messages for which no accumulator exists.
These lines are missing in org.apache.flink.table.runtime.operators.aggregate.GroupAggFunction:
The bug has not been verified. I spotted it only by looking at the code.