Need a plan of the KIngress prober for HTTPRoute

[nak3]

KIngress has prober function but HTTPRoute does not have it.
Additionally, the HTTPRoute's "admitted" status will not work as it is different from "Ready" status. Please see - https://kubernetes.slack.com/archives/CR0H13KGA/p1611865803004000

[markusthoemmes]

This is super important for the semantics of our Ingress, but I honestly have no clue on how to generically solve this.

We kind of have a same'ish problem in Openshift where we have a separate routing plane that gets configured for every Knative ingress. The networking setup of these routes can change from cluster to cluster and there is no guarantee that the knative-serving namespace can even reach those routers. There's also no generic way of listing all possible routers which might even be worse with IngressV2, seeing as there will be different implementations on it.

Ideas super welcome 😅

[howardjohn]

Previously all of the net-xyz controllers did their own probing, right? So the closest match to this would be to have all IngressV2 implementors do the same, by pushing for a true "Ready" status field in the API?

That seems like the most pure way to implement it - however, I have some pretty major concerns about implementors actually implementing it correctly for knative semantics. AFAIK no Envoy based implementations would be able to do this without a lot of efforts, which accounts for ~half of the currently implementors I think.

[nak3]

Previously all of the net-xyz controllers did their own probing, right?

Yes but strictly speaking, Knative has a common library for probing as:

https://github.com/knative/networking/blob/45fe1f5dd35eaebaf6f8ce03ab0983ede9497369/pkg/status/status.go#L135

and net-xyz controllers just calls it like:

(e.g. net-istio)
https://github.com/knative-sandbox/net-istio/blob/c4e24bc7750441150ef56e8ece9c937661993c9d/pkg/reconciler/ingress/controller.go#L131-L140
https://github.com/knative-sandbox/net-istio/blob/c4e24bc7750441150ef56e8ece9c937661993c9d/pkg/reconciler/ingress/ingress.go#L230-L233

So the closest match to this would be to have all IngressV2 implementors do the same, by pushing for a true "Ready" status field in the API?

Yes, that's correct. The perfect solution is that Ingress v2 supports Ready status in their API but I think they will not accept it due to the reason you mentioned.

[markusthoemmes]

@howardjohn I guess IngressV2 implementors could do the same probing we do in Knative to achieve the same effect. I'm aware that Envoy doesn't give you the "I"m actually ready to serve this config" signal via its config, which is a bummer. The difference is that the IngressV2 implementor intimitaly knows the router while Knative only knows the IngressV2 plane, which makes probing on our end brittle.

[howardjohn]

I'm aware that Envoy doesn't give you the "I"m actually ready to serve this config" signal via its config, which is a bummer

It does give you an ACK. In theory, this means it is ready to serve - I have not put the theory to test though.

I think it may be challenging for IngressV2 implementations to do probing since they need to handle arbitrary configs, whereas Knative has full control and a bounded set of possible routes defined, etc. I suppose you could probably always add some route not defined by the user with some special header or other match clause that is a NOP perhaps... but even then, there is no guarantee for all implementations that the control plane can even reach the data plane.

[markusthoemmes]

Right, exactly. That "not guaranteed to reach the data plane" is the problem we'll have in Knative too 😢

[github-actions]

This issue is stale because it has been open for 90 days with no
activity. It will automatically close after 30 more days of
inactivity. Reopen the issue with /reopen. Mark the issue as
fresh by adding the comment /remove-lifecycle stale.

[github-actions]

This issue is stale because it has been open for 90 days with no
activity. It will automatically close after 30 more days of
inactivity. Reopen the issue with /reopen. Mark the issue as
fresh by adding the comment /remove-lifecycle stale.