Proposal for routing to named instances

[kriskowal]

This captures an idea that @mranney communicated for enabling health checks to specific instances through Hyperbahn.

r @jsu1212 @jcorbin @Raynos

[prashantv]

I thought tags might be a nicer solution rather than doing something specifically for routing to an instance.

[jc-fireball]

We need to discuss more about that. This is also a problem I am trying to solve right now. The instance name for some python service are same among some instance. For example, the "service A" have 10 processes are naming "serviceA"

[anson627]

@junchaoWu interesting! How come different service instances have same name? aren't they named svc-0, svc-1, svc-2 etc.? I mean how deployment system or service controller would allow that.

[jc-fireball]

This is a legacy issue from python uwsgi world. Uwsgi spawns up multi-process of instances and all of them listen to same port.

[anson627]

thanks @junchaoWu! good to know.

[kriskowal]

@prashantv Yes, tags instead of instance name is the more general approach. You also mentioned using vector clocks. Let’s work on another iteration of this proposal. Each service would need to regenerate an index[tag]:set based on the most advanced ad/unad per worker.

[Raynos]

I like this proposal as is.

It's really simple and can be shipped.

We are missing only one thing. Let's change init request and init response to add instance_name as an optional header so that it's done at init time and not at advertise time.