feat(flowcontrol): Add ImmediateResponse abort mechanism for evicting in-flight requests

[RishabhSaini]

resolves "How to evict?" part of #2632

Integrate ext_proc ImmediateResponse as the abort mechanism for evicting dispatched requests from vLLM. When eviction triggers, closing the request's abort channel causes the Process() loop to send ImmediateResponse(503), making Envoy reset the upstream connection. vLLM detects the disconnect and frees KV blocks.

• AbortRegistry bridges eviction plugin and ext_proc Process() loop
• ImmediateResponseAborter closes abort channels via sync.Once
• recvOrAbort wraps srv.Recv() to select on abort signals
• Tests cover aborter, registry, recvOrAbort, plugin integration, and concurrent eviction+completion races

[k8s-ci-robot]

Hi @RishabhSaini. Thanks for your PR.

I'm waiting for a kubernetes-sigs member to verify that this patch is reasonable to test. If it is, they should reply with /ok-to-test on its own line. Until that is done, I will not automatically test new commits in this PR, but the usual testing commands by org members will still work.

Tip

We noticed you've done this a few times! Consider joining the org to skip this step and gain /lgtm and other bot rights. We recommend asking approvers on your previous PRs to sponsor you.

Once the patch is verified, the new status will be reflected by the ok-to-test label.

I understand the commands that are listed here.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[netlify]

✅ Deploy Preview for gateway-api-inference-extension ready!

Name	Link
🔨 Latest commit	a9cbf1c
🔍 Latest deploy log	https://app.netlify.com/projects/gateway-api-inference-extension/deploys/69dd45d1308fa4000868ada3
😎 Deploy Preview	https://deploy-preview-2737--gateway-api-inference-extension.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[nirrozenbaum]

I recommend being consistent with names.
eviction is consistent with evictor. aborter is yet another naming option but is not consistent with the existing code.

[ahg-g]

/ok-to-test

[ahg-g]

/assign @LukeAVanDrie

[ahg-g]

/test all

[LukeAVanDrie]

This spawns a new goroutine to call srv.Recv() every time it is invoked.

If the abortCh triggers first, this method returns errEvicted, but the spawned goroutine remains blocked on srv.Recv() until the stream is eventually closed or data arrives. While this leak is technically bounded by the lifetime of the gRPC stream (which closes shortly after we send the ImmediateResponse and return nil), it is still a sub-optimal pattern to spawn goroutines in a loop for every message received.

Consider starting a single reader goroutine for the lifetime of the stream that reads from srv.Recv() and pumps messages into a channel. You can then use Go's property where selecting on a nil channel blocks forever to dynamically enable listening to the abortCh only after it is populated (post-scheduling).

For example:

// At the top of Process()
recvCh := make(chan recvResult, 1)
go func() {
    for {
        req, err := srv.Recv()
        recvCh <- recvResult{req: req, err: err}
        if err != nil { return }
    }
}()

var abortCh chan struct{} // Initially nil

for {
    select {
    case result := <-recvCh:
        // handle message...
        // When scheduled: abortCh = s.abortLookup.Get(abortRequestID)
        
    case <-abortCh: // Blocks forever while nil
        // handle eviction
        return nil
    }
}

[RishabhSaini]

agreed, attempted to fix this using claude

[LukeAVanDrie]

This looks good, thanks! Added one suggestion, then this PR lgtm

[ahg-g]

This is a misuse of the pluggable framework. This logic represents the runtime that executes the eviction logic and it should not be modeled as a plugin. We need to put this guidance in a framework's dev guide: Plugins represent entities that a user can enable/disable via the config api, and they only reside under framework/epp/plugins.

In this case, there should be an evictor subsystem instance that gets passed to request control, for example PreRequest becomes evictor.TrackRequests.

We can do this clean up as a followup.

@kfswain @LukeAVanDrie

[ahg-g]

Created: llm-d/llm-d-inference-scheduler#818

[LukeAVanDrie]

/approve

/lgtm

[LukeAVanDrie]

/lgtm

/approve

[ahg-g]

/approve

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: ahg-g, LukeAVanDrie, RishabhSaini

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details

Needs approval from an approver in each of these files:

• OWNERS [ahg-g]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment