Avoiding thundering herd issue with aggregation window start/stop

[anvinjain]

When backend dies and comes back up, it receives all associated process groups and starts aggregation windows for all of them. They all compete for slots, some of them get it, rest wait till next aggregation window switch. Since all of them start at the same time, they will stop at the same time as well(provided aggregation window duration is same for all, which is true for v1 release at least). So slot competition will happen again and can lead to starvation for some process groups. Also contention on buffer pool when serializing profiles.

Possible solution is to introduce chaos(jitter) when starting profiles so that they space out well.