Back to engineering

4ROUTER / Serverless

Designing for cold starts

Understanding the path from launch to readiness.

Engineering note

Startup has multiple stages

From placement to a usable endpoint, startup may include image preparation, process launch, weight reads, device loading and warmup. Recording stages separately reveals where the wait occurs. Process creation is not a substitute for readiness.

Caches and warm capacity have costs

Local weight caches reduce repeated downloads but need capacity and version management. Warm replicas reduce startup waits while retaining resources. Choose a policy against request frequency, model changes and acceptable waiting time.

Make failures recoverable

Corrupt weights, load failures and resource shortages need distinct errors and recovery paths. Do not admit traffic before readiness. During scaling and rolling updates, inspect what callers experience while waiting, not just the eventual successful state.

Before implementation

  • Startup-stage records
  • Weight-cache versions
  • Readiness and retry policy

Further reading

Kubernetes · Container probes

4ROUTER / SYSTEMS RESEARCH

Mechanisms and experiment design

Understand how operators, KV state and scheduling shape an inference path.

Read the systems notes in Resources
GEMM / KERNELS

From operators to serving: cuBLAS & cuDNN

Identify the limiting resource before selecting the implementation.

STATE / CAPACITY

KV cache: from tensor shapes to capacity budgets

Context length is a state budget that grows with concurrency.

TIERS / TRANSFER

LMCache: placing KV state in a storage hierarchy

Avoiding prefill requires paying for state access.

Let’s talk infrastructure.

Connect to the platform, or discuss models, performance and deployment.