Back to engineering

4ROUTER / Observability

Following a request end to end

Connecting admission, execution and accounting.

Engineering note

Give the request a shared identity

A request crossing gateway, model service and records needs correlation. Request identifiers and trace context serve related but distinct purposes. Keep their propagation clear across service boundaries so timestamps refer to the same work.

Record stages, not unnecessary content

Trace spans describe units of work and their timing relationships. Useful stages include queuing, selection, upstream calls and record writes. Do not copy prompts, responses or credentials into telemetry simply for convenience. Collection should match access and retention policies.

Move from a symptom to its path

For a slow request, locate the delayed stage before interpreting errors and resources. Cost discrepancies require attempts and actual usage, not just the final response. Sampled tracing does not replace business records that require complete retention.

Before implementation

  • Correlation propagation
  • Sensitive-content boundaries
  • Retention and access control

Further reading

OpenTelemetry · Traces

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.