Bypass supervisor's nasty callbacks stack and make it quick!
Bypass supervisor's nasty callbacks stack and make it quick!
$ pip install supervisor-quick
And add the following config to supervisord.conf
.
[ctlplugin:quick]
supervisor.ctl_factory = supervisor_quick:make_quick_controllerplugin
Then start supervisorctl
and use quickstart
, quickstop
and
quickrestart
to start/stop/restart processes.
> quickstart app:0
> quickstart app:
> quickstart ap*
> quickstart all
> quickstop app:1
> quickstop app:
> quickstop ap*
> quickstop all
> quickrestart app:2
> quickrestart app:
> quickrestart ap*
> quickrestart all
It effects supervisorctl
, so you don't have to restart the whole
supervisord to make it work.
I write this plugin because supervisor is just tooooo slow in start/stop/restart app server in our prod servers.
And I checked the source code and found it is because of the
nasty callbacks stack, and this is a quote from source code
supervisor/rpcinterface.py
.
XXX the above implementation has a weakness inasmuch as the first call into each individual process callback will always return NOT_DONE_YET, so they need to be called twice. The symptom of this is that calling this method causes the client to block for much longer than it actually requires to kill all of the running processes. After the first call to the killit callback, the process is actually dead, but the above killall method processes the callbacks one at a time during the select loop, which, because there is no output from child processes after e.g. stopAllProcesses is called, is not busy, so hits the timeout for each callback. I attempted to make this better, but the only way to make it better assumes totally synchronous reaping of child processes, which requires infrastructure changes to supervisord that are scary at the moment as it could take a while to pin down all of the platform differences and might require a C extension to the Python signal module to allow the setting of ignore flags to signals.
And this plugin will do a quick
start/stop/restart action that bypass
all the callback checks, making it lightning fast.
It also have wildcard concurrent execution support, keeping it fast regardless of processes amount. (This function is inspired by supervisor-wildcards)
An example time demo for a app server with numprocs set to 32 to show how quick
supervisor can be with quick
command.
$ supervisorctl status
app:0 STOPPED
app:1 STOPPED
app:10 STOPPED
......
app:7 STOPPED
app:8 STOPPED
app:9 STOPPED
$ time supervisorctl start app:
24: started
25: started
26: started
......
18: started
31: started
30: started
supervisorctl start app: 0.06s user 0.02s system 0% cpu 48.442 total
$ time supervisorctl stop app:
24: stopped
25: stopped
26: stopped
......
18: stopped
31: stopped
30: stopped
supervisorctl stop app: 0.06s user 0.03s system 0% cpu 36.278 total
$ time supervisorctl quickstart app:
app:25: started
app:24: started
app:27: started
......
app:1: started
app:8: started
app:9: started
supervisorctl quickstart app: 0.09s user 0.03s system 19% cpu 0.618 total
$ time supervisorctl quickstop app:
app:26: stoped
app:27: stoped
app:22: stoped
......
app:0: stoped
app:9: stoped
app:8: stoped
supervisorctl quickstop app: 0.09s user 0.04s system 68% cpu 0.196 total