Skip to content

LLM infrastructure

Request counts hide much of the work in an LLM platform. Build a model of serving capacity, admission and tenant isolation before choosing routing or retry policies.

Start with serving mechanics

  1. LLM Inference Is a Scheduler Problem

    See how prefill, decode, continuous batching and KV memory determine what an inference engine can safely run.

    Essay

Measure load and latency

  1. Your LLM Dashboard Is Lying About Load

    Replace request-count intuition with signals that reveal token, memory, queue and scheduler pressure.

    Essay

  2. Your LLM System Isnt Slow. Your Tail Is Slow.

    Separate time to first token, stream pauses and queue delay to find where users actually wait.

    Essay

Control admission and capacity

  1. Your LLM Rate Limiter Is Counting the Wrong Thing

    Reserve for uncertain token work before it reaches the backend, then reconcile actual use.

    Essay

  2. One Tenant Can Break Your LLM Platform Without Sending Many Requests

    Design fair queues, KV budgets and isolation for workloads whose request counts hide their resource use.

    Essay

Route and recover deliberately

  1. Your LLM Router Can Make Cheap Models Expensive

    Choose routes across a session using cache locality, quality, provider health and switching cost, not token price alone.

    Essay

  2. Retries Make LLM Systems Less Reliable

    Bound retries and fallback so provider failures do not amplify load, cost or behavior changes.

    Essay

Browse all matching essays

Keep reading

Continue with a guided sequence of free production engineering essays.

Find your next reading path