vdbergh
vdbergh
One may wonder however what the point is of this summing of WLD data corresponding to different sets of parameters...
I am just seeing this. > with the only remaining major issue that `nElo` has wrong assumptions about the draw distribution. This is incorrect. `nElo` does not make _any_ assumptions...
I don't understand. nElo also uses game pairs (it is computed from the pentanomial frequencies).
@vondele Can you explain why the internal throttling in Fishtest does nothing (at most five concurrent accesses to `request_task`)? I had the PR #1987 to set the number to 1...
> PROD running with 10 threads per instance. I will check in the log for request rejected. If Disservin is correct then the # of threads is irrelevant because of...
I think the number of connections in waitress should be equal to the number of threads (so that waitress does not queue connections) and then the value of the request_task...
The thing is that 100 is too big. Waitress will simply queue connections without giving Fishtest a chance to reject them. That's why we are seeing 30s for api calls....
Yes nginx caps the connection time. But either waitress or Fishtest should reject the connections. It should not be left to nginx.
> It's really the job of nginx. We can set limits for host, uri, even for any single path. > > ```nginx > limit_conn_zone $host zone=perhost:10m; > limit_conn perhost 10;...
Note that Fishtest will reject the connection immediately. There will be no 30s timeout.