Start at the boundary.
A field manual for the space between a client action and a dependable outcome. Seven guides. Explicit assumptions. Failures you can reproduce.
Building a service
Map the request, decide who owns the effect, then make recovery safe.
Lifecycle → idempotency ↗Testing a failure
Separate a lost response from a lost effect. Compare the resulting traces.
Callbacks → the lab ↗Operating a platform
Publish coherent content and investigate failures with less sensitive data.
Catalogs → observability ↗The manual
Expanded 17 November 2026- 017 min read
The request lifecycle
Follow a session request across service boundaries, locate ambiguous outcomes, and design a recovery path.
- 027 min read
Callbacks, retries & recovery
Separate delivery from processing. Design bounded retries, durable acknowledgement and a useful dead-letter workflow.
- 037 min read
One event. One intended effect.
Put idempotency at the commit boundary, handle concurrent duplicates, and understand what exactly-once claims leave out.
- 047 min read
A catalog is a moving snapshot
Synchronize paginated content feeds with stable identities, staged versions, explicit deletions and deliberate locale fallbacks.
- 057 min read
Follow the signal, not the user
Connect requests, attempts and outcomes with correlation IDs, useful alerts and privacy-aware telemetry.
- 066 min read
Arrival is not application order
Follow a two-event stream through a deliberate buffer policy. Distinguish event identity, application order and the evidence left by a lost acknowledgement.
- 076 min read
Design the evidence before the failure
Build a six-scenario acceptance matrix with explicit oracles, bounded replay and privacy-minimal traces. Learn what a passing simulation cannot establish.
Read the assumptions first. These are original, synthetic designs. A concrete provider may use a different contract, retry window or consistency model. Each guide links the sources behind its factual claims.