refactor the control loop and handle sigterm #65
Conversation
| @@ -184,10 +180,6 @@ def send(self, value): | |||
| self._write_packet(data) | |||
| return len(data) + self._packet_len.size | |||
|
|
|||
| def recv(self, timeout=None): | |||
| packet = self.reader_queue.get(block=True, timeout=timeout) | |||
| return packet | |||
There was a problem hiding this comment.
got rid of the reader queue here and instead just invoke a callback letting the owner of the pipe handle packets how they wish (using a deque, notify the control pipe, etc)
|
|
||
| def _signal_sighup(self, signum, frame): | ||
| self.logger.info('Received SIGHUP, triggering a reload.') | ||
| self.monitor.set_changed() |
There was a problem hiding this comment.
this old version was acquiring locks in the handler - now we just write to the control pipe
| @@ -61,7 +61,7 @@ def is_alive(self): | |||
|
|
|||
| def stop(self): | |||
| if self.is_alive(): | |||
| self.process.terminate() | |||
| self.process.kill() | |||
There was a problem hiding this comment.
hupper now catches SIGTERM so instead we want to just SIGKILL in the test suite once we're done with the test
| worker.wait(shutdown_interval) | ||
| elif signal == ControlSignal.SIGCHLD: | ||
| if not worker.is_alive(): | ||
| break |
There was a problem hiding this comment.
currently on windows SIGCHLD will not occur, but it's not a big deal because we have an open pipe with the child. When the child dies, the pipe is closed and we emit a WORKER_COMMAND with a None packet and we're able to exit pretty quickly via the if cmd is None check above.
mmerickel
force-pushed
the
handle-sigterm
branch
from
bfcf284 to
1b68c43
Compare
mmerickel
force-pushed
the
handle-sigterm
branch
from
102b4c0 to
4294fd2
Compare
mmerickel
force-pushed
the
handle-sigterm
branch
from
4294fd2 to
cffa460
Compare
mmerickel
force-pushed
the
handle-sigterm
branch
from
cffa460 to
c27df8f
Compare
mmerickel
force-pushed
the
handle-sigterm
branch
from
8677f75 to
fbf8ab4
Compare
This way we make sure that when the other thread goes to read self.monitor.is_changed it is actually set, especially if for some reason sending the "signal" down the pipe causes the read to wakeup before the thread finishes processing.
This reverts commit fbf8ab4.
fixes #64
fixes #63