#42345 removed some indexes because they create too many deadlocks when running many concurrent consumers.
Yet, as reported in the linked issue, those indexes are useful when processing large queues (typically the failed messages queue).

#45007 is proposing to add an option to force the indexes back, but I don't like it because it requires ppl to learn about the issue. I looked for a more seamless solution and here is my proposal.

Instead of possibly triggering the deadlock duringack()/reject(), I propose to use a soft-delete there, and do the real delete inget(). This makes ack/reject safe because they don't alter any indexes anymore (the delete was), and this moves deadlock exceptions to the same function that creates the locks. This allows the retry mechanism inDoctrineReceiver to recover from at most 3 consecutive deadlock exceptions. There can be more, and in this case, the consumer will stop. But this should be much less likely.

(yes, I did create a reproducer to work on this issue ;) )