Draft page

This page is an outline. It describes what will be covered and is not yet complete technical documentation.

A generic load balancer treats inference replicas as interchangeable and requests as equal cost. Neither holds: replicas differ in cached prefixes and queue depth, and request cost varies by orders of magnitude.

Inference-aware routing uses those signals, and typically improves both latency and cache hit rates.

What you will learn

  • Why generic load balancing underperforms for inference traffic.
  • Which signals a router can use to place a request.
  • How prefix-aware routing raises cache hit rates.
  • How routing interacts with autoscaling.

This page is an outline. The subsections below are the planned structure; they are filled in as the handbook is written.

Why round-robin underperforms

Draft

Useful routing signals

Draft

Queue-aware routing

Draft

Prefix-aware routing

Draft

Multi-model routing

Draft

Interaction with autoscaling

Draft