feat(BA-992): Offload health check capability to AppProxy

[kyujin-cho]

Resolves #3051 (BA-992).

Summary

• Refactor model service health check architecture to offload health checking to AppProxy
• Update API specifications for endpoint route management and health check configuration
• Implement separate model entity handling for health check configuration
• Add support for reading health check information from model-definition.yaml

Breaking Changes

Important: This PR breaks health check capability on OSS AppProxy. Future work will restore support, but for now disable the health check feature in model-definition.yaml to use model service on Open Source Backend.AI.

Changes

API Specifications

• Updated model serving event types with new ModelServiceStatusEventArgs base class
• Added EndpointRouteListUpdatedEvent for endpoint route synchronization
• Refactored anycast/broadcast event structures for better separation of concerns

Health Check System

• Added ModelServiceHelper class replacing ModelServicePredicateChecker
• Implemented health check configuration reading from model-definition.yaml
• Added health check configuration passing to AppProxy during endpoint creation
• Separated model definition validation into discrete functions

Route Management

• Updated endpoint route generation to use Redis for AppProxy communication
• Modified session lifecycle handling for route status updates
• Added generate_redis_route_info method for serializable connection data
• Implemented notify_endpoint_route_update_to_appproxy for real-time updates

Database Changes

• Enhanced endpoint model with Redis route info generation for AppProxy integration
• Updated routing status management in session callbacks
• Improved error handling and retry logic for route operations

Technical Details

The health check system now reads configuration from either:

1. Runtime variant profiles for predefined endpoints
2. model-definition.yaml for custom runtime variants

Health check configuration is passed to AppProxy during endpoint creation, allowing AppProxy to handle health checking independently. Route connection information is stored in Redis for AppProxy consumption with the key pattern:
endpoint.{endpoint_id}.route_connection_info

Testing

Existing tests should continue to pass. New functionality requires model service integration testing with AppProxy components.

[Copilot]

Pull Request Overview

This PR refactors how health checks and route updates are offloaded to AppProxy. Key changes include:

• Introducing ModelServiceHelper and reading health check config from model-definition.yaml
• Switching route updates to Redis-based notifications (valkey_live) instead of direct AppProxy HTTP calls
• Adding Pydantic HealthCheckConfig and YAML-driven health check extraction

Reviewed Changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/ai/backend/manager/services/processors.py	Injected ValkeyLiveClient into ServiceArgs
src/ai/backend/manager/services/model_serving/types.py	Added RouteConnectionInfo dataclass
src/ai/backend/manager/services/model_serving/services/model_serving.py	Refactored endpoint creation and route update flows to use ModelServiceHelper and Redis
src/ai/backend/manager/registry.py	Implemented AppProxy endpoint creation, health-check support, and Redis notifications
src/ai/backend/common/data/config/types.py	Added HealthCheckConfig Pydantic model

Comments suppressed due to low confidence (3)

src/ai/backend/manager/registry.py:3586

• Add unit tests for get_health_check_info to cover both standard runtime variants and CUSTOM model-definition.yaml parsing to ensure health check configurations are correctly extracted.

    async def get_health_check_info(

src/ai/backend/manager/registry.py:3640

• [nitpick] Add or update the docstring for create_appproxy_endpoint to explain parameters, return value, and health_check payload format for better maintainability.

            async with session.post(

src/ai/backend/manager/registry.py:3658

• The create_appproxy_endpoint payload no longer includes any 'apps' or inference routing info; ensure AppProxy can derive routing solely from Redis notifications or re-add necessary fields to the payload if required.

                    "health_check": health_check_information.model_dump(mode="json")

[Copilot]

[nitpick] The RouteConnectionInfo dataclass is defined but never used; consider removing it or integrating it into generate_redis_route_info for consistent typing.

Suggested change

	@dataclass
	class RouteConnectionInfo:
	app: str
	kernel_host: str
	kernel_port: int
	# Removed the unused RouteConnectionInfo dataclass as it is not referenced anywhere in the code.

[Copilot]

Consider wrapping Redis operations in notify_endpoint_route_update_to_appproxy within try/except to handle potential valkey_live errors and avoid unhandled exceptions disrupting routing updates.

[HyeockJinKim]

RouteLifecycleEvent?

[HyeockJinKim]

Please fix this conflict tomorrow. @kyujin-cho

[HyeockJinKim]

Please resolve conflicts.