Problem
A message that a consumer processed successfully, and reported as processed, was redelivered about 12 seconds later, against a 60-second visibility timeout and with a single consumer running. The redelivery reported attempt: 2, so it came through the visibility-timeout path on a delivery-state row that was still unacked.
Impact
The duplicate propagated: Runway answered the same merge-conflict-check twice, and the batch stage created a second batch for the same request. The pipeline absorbed it, and the duplicate batches failed, but each one cost a full speculation and build cycle. If a successful ack can be redelivered at a fifth of its visibility timeout, at-least-once delivery is firing far more often than intended.
Not yet established
Either the ack did not set acked, despite the consumer reporting success, or invisible_until was shorter than configured. The delivery-state row had already been garbage-collected, so it could not be inspected. This reproduced under load (a 20-request demo run) and not on smaller runs.
Status
No fixing change has been identified on main.
Visitar URL original
Problem
A message that a consumer processed successfully, and reported as processed, was redelivered about 12 seconds later, against a 60-second visibility timeout and with a single consumer running. The redelivery reported
attempt: 2, so it came through the visibility-timeout path on a delivery-state row that was still unacked.Impact
The duplicate propagated: Runway answered the same
merge-conflict-checktwice, and the batch stage created a second batch for the same request. The pipeline absorbed it, and the duplicate batches failed, but each one cost a full speculation and build cycle. If a successful ack can be redelivered at a fifth of its visibility timeout, at-least-once delivery is firing far more often than intended.Not yet established
Either the ack did not set
acked, despite the consumer reporting success, orinvisible_untilwas shorter than configured. The delivery-state row had already been garbage-collected, so it could not be inspected. This reproduced under load (a 20-request demo run) and not on smaller runs.Status
No fixing change has been identified on
main.