More resiliant approach for in-flight requests on terminating Pods about traefik HOT 5 OPEN

Hello @alen-z and thanks for your interest in Traefik,

This is an interesting idea, there is already a pull request #10664 which brings the EndpointSlices support which might fix this issue.

@jnoordsij wdyt?

from traefik.

I've looked into the mentioned KEP, but I just don't see how exactly it should benefit Traefik. In both the existing handling of endpoints and the one I implemented with the EndpointSlice API in #10664, any terminating endpoints should already be considered invalid as they should be marked as not ready by Kubernetes.

The preStop hook will help cover the possible time overlap between Traefik being aware of the endpoint no longer be available (whether through Endpoint or EndpointSlice API) and the pod being actually terminated; this is described at various places, see e.g. https://itnext.io/how-do-you-gracefully-shut-down-pods-in-kubernetes-fb19f617cd67. AFAICS, the mentioned proposal does not offer any alternative to that.

What could be done after #10664 is creating some kind of configuration flag (or even/later default behavior), that does broadly the same thing as the mentioned KEP: rather than relying on the ready flag, Traefik could allow all serving=true endpoints and then using only those with terminating=true if there are no endpoints with terminating=false, to have a "theoretical" better chance of finding suitable endpoints.

However, I'm not sure if this has any real-world benefit, as the motivation mentions the problem at hand being "When using Service Type=LoadBalancer w/ externalTrafficPolicy=Local" while Traefik should typically be sending traffic to ClusterIP (or NodePort) services. Moreover the Non-Goals lists "Handling terminating endpoints for other consumers of the EndpointSlice API, such as ingress controllers or external load balancers.", leading me to believe this is not relevant for Traefik?

from traefik.

Hey @jnoordsij, great gist and appreciate the time invested. Also, pretty nice to see #10664. I wasn't aware, pretty new from the Traefik kitchen.

Ultimate goal: Removing the need to have preStop hook and no new requests sent to terminating Pod + make a Pod stay alive to serve in-flight requests (automated preStop bacisally). This means finding the way to remove this gap between Traefik being notified and Pod being terminated (if preStop does not exist).

I have one thing in mind: Can Traefik intervene in the process of EndpointSlice and the state that controls endpoint termination? Basically, have Traefik as one of the validation gates before changing the state for termination? This puts Traefik in a position to acknowledge and influence the termination action. Example: Pod would not terminate until Traefik responds that it has no in-flight request — then EndpointSlice state changes. Or maybe Traefik can influence Pod lifecycle itself to say when it's ready to allow Pod to terminate, finishing in-flight requests even before Pod started the termination?

How to involve Traefik in a similar way that kube-proxy is involved is a good question. I'd need to look more into it to propose implementation details, but from the top of my head: finalizers, maybe. Add Traefik finalizer, remove when ready to terminate after finishing in-flight. Or Admission Controller of some sort, though I'm not aware if relevant events to EndpointSlice resource are passing there?

from traefik.

I see! I do like the idea, although I think it might be very challenging to achieve such a thing in practice.

But as far as I can see, that goal does not have any direct relation to KEP-1669 or my PR #10664, given they're really about something else (namely the endpoint part) and do not actively attempt to alter Pod termination logic itself in any way.

from traefik.

Yes, you are right. KEP was just initial spark that evolved.

Glad you find it interesting. If there is a way to try this, maybe we can put it behind a flag to start with.

from traefik.