[FEATURE REQUEST] Customization of behavior when rr is unable to allocate worker

[jjsaunier]

👋 @48d90782

follow up after discord discussion

Goal : Distinct error from worker allocation to be able to retry on another rr instance at load balancer level.

Current State : When rr is not able to allocate a worker, it's send a 500, which is not the best option because application that rr serve could also send 500 HTTP code.

IHMO, rr should give the possibilty to the user to:

• send a 504 - Gateway Timeout (recoverable)
• send a 502 Bad Gateway (not recoverable)
• Close the connection without responding (In my mind, traefik retry middleware work in this way) - https://doc.traefik.io/traefik/middlewares/retry/#:~:text=The%20Retry%20middleware%20is%20in,regardless%20of%20the%20response%20status.

With this 3 possibilities I think we good, if you see / have another use case, don't hesitate

https://github.com/spiral/roadrunner/blob/5dc83ef5eee77f4d1ef557e5f8b566e75892680d/service/http/handler.go#L134-L151

[rustatian]

👋 @jjsaunier
I quite agree with your proposal. I'll think about implementation because there is not only an "allocate worker" error that can be returned from the pool.Exec method. There are also fmt.Errorf("payload can not be empty"), fmt.Errorf("worker is not ready (%s)", w.state.String()) and may be more.
I think that it is better to distinct all RR errors from the errors which come from the PHP.
As for the fast solution, that you can do on your own - is to temporarily replace w.WriteHeader(500) with w.WriteHeader(502) or 504 and then build docker image (docker build -t roadrunner:local -f Dockerfile .) or invoke make build target in the Makefile.

[jjsaunier]

Indeed, i'm only experiencing the issue with allocate worker, so no experience on payload can not be empty, worker is not ready but sounds good to dissociate all rr error vs app error 👍

Thanks for the quick solution, unfortunately traefik (which I'm using) does not support retry based on status code at this time... (I hope a day it will!)

So currently my trick to let traefik retrying, I close the connection without responding :

hj, ok := w.(http.Hijacker)

if !ok {
	http.Error(w, "", http.StatusGatewayTimeout)
	return
}
		
conn, _, err := hj.Hijack()
conn.Close()

[rustatian]

@jjsaunier #472
1.9.2 milestone

[jjsaunier]

You rocks @48d90782! 🥇

I left comment on pr

[rustatian]

@jjsaunier https://github.com/spiral/roadrunner/releases/tag/v1.9.2

[wolfy-j]

I believe we will need to add the same thing for RR2 down the road.