Shuffle dequeue #63
Description
If there's a sequence of tasks, which are particularly fast, for example, because they fail or because their workload is exceptionally small, this leads to a surge in compute tasks being completed, which produces spikes in natural traffic and latency, potentially disrupting the workers. This is best mitigated by shuffling the sequence of the jobs being de-cute, which also helps with estimating the remaining runtime of the tag.