[Serve][3/5] Add ingress request router targets to TargetGroup

[eicherseiji]

Why this is needed

PR2 lets an app identify a peer ingress request router. This PR carries that role through the controller/runtime path so HAProxy can send /internal/route lookups to router replicas while forwarding data-plane traffic to the normal ingress replicas.

Summary

• Propagates ingress_request_router through deploy args, DeploymentInfo, application state, and replica startup.
• Adds TargetGroup.ingress_request_router_targets so HTTP target groups can expose router replicas separately from backend replicas.
• Allows router replicas to own direct-ingress HTTP ports even though they are not the app ingress deployment.

Stack: #62667 -> #62680 -> this -> #62669 -> #62670

Test plan

• Added / updated unit coverage in python/ray/serve/tests/unit/test_application_state.py
• Added controller target-group coverage in python/ray/serve/tests/unit/test_controller_direct_ingress.py
• python -m py_compile on the edited files
• git diff --check
• Repo commit hooks / linters on commit

[gemini-code-assist]

Code Review

This pull request implements an "ingress bypass" feature in Ray Serve, allowing direct data plane traffic to replicas via custom ASGI apps while utilizing a dedicated router deployment for Lua-based routing decisions. The changes include updates to the controller's target group management, the addition of a detached registry for direct ingress endpoints, and new replica-side logic for starting HTTP servers for custom apps. The review feedback identifies several critical issues: performance degradation caused by blocking the event loop with ray.get in asynchronous methods, a potential memory leak in the _DirectIngressRegistry due to the lack of an unregistration mechanism, missing imports for errno and retry constants, and the inclusion of unused helper functions in the controller.

[kouroshHakha]

Nice incremental wire-up — most of the role plumbing was done in #62680, so this PR is appropriately small. The router replicas correctly receive direct-ingress HTTP ports without gRPC, and the new ingress_request_router_targets field gives HAProxy what it needs without disturbing the data-plane target list.

A few things to consider before merge:

1. Validation in _set_target_state raises after partially clearing the cached ingress names — minor but leaves the object in a half-mutated state on exception.
2. The "no replicas → return []" short-circuit means routers won't be advertised until ingress is also up. Probably intentional, but worth a one-line comment so future maintainers don't try to "fix" it.
3. Naming nits inline: a few ingress_request_router / _ingress symbols would read better as is_*.
4. Stale gemini/cursor bot comments above no longer apply — feel free to dismiss.

No critical findings.

Note

This review was co-written with AI assistance (Claude Code).

[kouroshHakha]

All four points from my prior review are addressed cleanly:

1. _set_target_state now builds target_state first and only assigns self._ingress_*_deployment_name / self._target_state after — no more half-mutated state if construction raises. New test_application_state_clears_stale_ingress_request_router covers the clearing-on-transition case.
2. The "no ingress replicas → return []" short-circuit is now commented as load-bearing.
3. Replica/ReplicaActor parameter and attr renamed to is_ingress_request_router, consistent with is_ingress.
4. Bonus: gRPC config fetch skipped on router replicas, HTTP-only nature of ingress_request_router_targets documented in both schema and controller docstring, and get_ingress_replicas_info docstring expanded.

LGTM.

Note

This review was co-written with AI assistance (Claude Code).