When under heavy load (5000-7000req/s) messages that are sent by workers don't reach the master instance anymore.
Tested on Windows 7 x64:
As stated here there is a hard limit that seems to cause the problem because especially on windows the message queue isnt flushed as fast as on linux based systems.
Yes, I could reproduce the issue. A fix has been landed in cfa2869 which avoids dropping messages. However there still is no good story for throttling send() calls at the moment.
seems we're past this point.