Cross posting from {T387032}

In T387032#10586579, @jcrespo wrote:

As an additional thought- I believe one of the previous mitigation was to enable the query killer on parsercaches as an additional safe step towards preventing overload. Now I think this should be reviewed as something that should be maintained, as I believe I saw a lot of rollbacks on processlist when shutting down some of the hosts (it normally only kills SELECTS, but it starts killing all queries when it goes over a higher limit of total connections). It would be nice to check if that was the origin of the ROLLBACKs, and while it wouldn't be the original cause of the overload, it could had made a bad thing into a worse thing.

So after overload cause is identified, it would be interesting to reevaluate if the query killers should be still enabled on parsercache to minimize write killings.