On Wed, Apr 27, 2016 at 9:30 PM, William A Rowe Jr <wr...@rowe-clan.net> wrote:
> On Wed, Apr 27, 2016 at 6:16 PM, Yann Ylavic <ylavic....@gmail.com> wrote: > >> >> So now, either we choose to not report workers' activity in between >> requests (there are just multi-purpose workers/request-pollers after >> all), or we save some "last request" informations in each conn_rec and >> restore them on the worker (score) handling the connection at each >> time (this is more reporting the workers' activity than the requests >> lines, which will pass from one worker to the other, even possibly >> duplicating requests lines with low activity). >> > > If that slot does other work, it should be recorded. But while that slot > sits unused, the last activity for that score slot must be preserved. > Just to clarify the event case... our server is not free-threaded. Once we are in a request, we are threadbound. The following request on that same worker is likely to jump threads, and it's fine, my brief testing today seems to confirm, but I presented the patch to the followers of PR 59333 for further review and testing. This will likely change in 3.0 (no, we aren't going to make requests free-threaded without bumping the version major - that's too drastic a behavior change.) Once a *request* is jumping from thread to thread, as opposed to the underlying connection, we are going to have to rethink how the scoreboard presents that information and how it is refreshed.