On 9/18/20 4:04 PM, Yann Ylavic wrote:
> This was lately reported to me (privately) with the root cause being
> r1819855 ([1], present in 2.4.x too) setting workers_were_busy=1 when
> connections_above_limit() is reached.
>
> Before r1819855, the #connections limit (based on
> AsyncRequestWorkerFactor) only caused listening sockets to be
> "paused", while now it also causes existing keepalive connections to
> be early terminated (see [2]), while the docs for
> AsyncRequestWorkerFactor state that it only happens when "all workers
> are busy".
>
> Since (I think) there is no point in killing connections before it's
> really needed (workers exhausted), I'm about to restore the original
> behaviour with the attached patch, WDYT?
>
Sounds sensible.
Regards
Rüdiger
