Spinning up a dedicated fly machine for a user?

[peterseibel]

I’m trying to design an architecture for a system where each user gets a dedicated fly machine for as long as they’re using it. is fly-replay the only way to establish affinity between a client and a specific fly machine? If so, am I wrong to think it’s unfortunate that every request (on overage) is going to have to hit two machines to get a response—the first one to check a cookie and reply with a fly-replay header which will then cause the correct machine to be given the request? Is there really no way for the original request to specify the specific machine that should handle it?

[rubys]

@peterseibel

Is there really no way for the original request to specify the specific machine that should handle it?

There is a way:

@jerome

We’ve now added a functionality where if you add fly-force-instance-id header to your request, with a value containing an instance id, we will forward the request to that instance only. This should also work with websockets and upgrades if the original request was made with the aforementioned header.

This is a bit of a footgun: we will not try other instances when you provide that header. If the instance is down, the connection will probably hang for a little bit and then close.

[peterseibel]

Is that even still a thing? I found that post before I asked my question but couldn’t find any mention of it in the docs or even in Google other than pointing to old posts in this forum.

If it is a thing, is it documented somewhere that I just missed?

[rubys]

I just tried it with an app of mine, sending requests to specific machine instances using curl; and I verified the correct routing in my logs. As it worked, I added some minimal docs:

• Dynamic Request Routing · Fly Docs
• Public Network Services · Fly Docs

[peterseibel]

Awesome! Thank you.

Maybe one last thing to add to the docs is what, if anything, we can count on happening if the instance is gone. Like, do we get a 502 Bad Gateway or some 500 at least? Or does it just hang and/or eventually timeout with no response?