Serial restart for a watcher's processes

[lucian1900]

It would be great if I could tell circusd to restart one process at a time within a watcher and not kill them all in parallel and then restart them all in parallel. For now, I'm working around this by parsing circusctl stats watchername and killing the processes one by one, expecting circusd to restart them.

The motivation for this is using circusd for a project which has:

• several HTTP workers listening on the same socket, putting tasks on a queue
• several queue workers that pop from the queue, process and return a result through the queue

I love that circusd lets me restart even the HTTP workers without downtime, but it would be nice to not have "slowtime" either, where during a watcher restart clients wait on a new worker coming up.

[Natim]

Hello, there is something in salt that lets you choose the batch-size (http://docs.saltstack.com/topics/targeting/batch.html) we could integrate something like that to the restart command.

[lucian1900]

Yeah, circusctl restart --batch=1 would be equivalent to my workaround. I would probably also call it through the API, which I assume Salt would be doing as well.

[tarekziade]

:+1:

[pomarec]

Shoud you close this issue ?

[mitsuhiko]

Why was this closed?

[pomarec]

I don't even remember why I was interested in this topic.

[k4nar]

I don't know, I don't see any similar option in Circus. I guess it would not hard to do something like salt, with --batch=1 or --batch=25%.