feature: enhance observability with instance-level and routing-level metrics

Describe the feature

Currently, the observability in production-stack is limited, primarily offering coarse-grained cluster-level metrics. Compared to the existing metrics by vLLM's observability system, they lack instance-level metrics and don’t provide sufficient insights for routing decisions—especially as we scale to support more complex or user-defined routing algorithms (e.g., like those in AIBrix).

I propose:

1. Reusing vLLM’s observability framework for instance-level metrics (e.g., per-instance latency, throughput), as it offers finer granularity and better usability.
2. Adding routing-specific metrics, such as: per-instance QPS (critical for current routing), routing decision metadata (e.g., algorithm used, fallback triggers, retries) for future complex routing logic.

Why do you need this feature?

• Instance-level: vLLM’s existing metrics are more detailed making them a better fit in production scenarios.
• Routing-level: As routing grows more complex (e.g., dynamic load balancing, kvcache-aware routing), observability will be key to debugging and optimization. Without it, diagnosing issues like skewed load or suboptimal routing becomes opaque.

Additional context

This is a demo based on vLLM's observability framework, implementing a session-based router with integrated real-time QPS monitoring for each vLLM instance at the routing-level.

Are there any other essential metrics I might have overlooked for effective monitoring? Please don't hesitate to provide input!
@YuhanLiu11 @ApostaC @Shaoting-Feng

Start by reviewing vLLM’s observability framework and production-stack’s existing cluster-level metrics. Compare the session-based router demo and its real-time per-instance QPS monitoring with the requested instance-level and routing-level metrics, then clarify the metric scope and validation approach with maintainers.