Priority: P1 (HIGH)
Each pattern buys one guarantee and charges operational complexity. Adopt on evidence, not on architecture fashion.
Selection Table
| Constraint observed | Pattern | Cost accepted |
|---|---|---|
| A write must update the database and publish an event atomically | Transactional outbox | A relay (poller or CDC) to run and monitor |
| Read shape and write shape have diverged and one model serves neither | CQRS | Two models, sync lag, doubled tests |
| The sequence of changes is itself the product (audit, replay, temporal queries) | Event sourcing | Projections, event versioning, no easy delete |
| A business transaction spans services and cannot hold a lock | Saga with compensations | Compensating action per step; no isolation |
| Exactly one instance may run a job or own a resource | Leader election (lease-based) | Fencing tokens; split-brain handling |
| Cross-cutting concerns repeated in every service | Sidecar or ambassador | Another process per pod to operate |
| A legacy or third-party model would infect the domain | Anti-corruption layer | A translation layer to maintain |
| Each client surface needs a different aggregate shape | Backends-for-frontends | One backend per surface, owned by that team |
| A monolith must be replaced without a rewrite freeze | Strangler fig | Dual routing and a live seam until the last slice moves |
Consistency Across a Boundary
- Never dual-write. Write to your own store, then publish from the outbox or CDC stream; anything else drifts silently.
- Saga orchestration puts the sequence in one coordinator: use it beyond ~4 steps, or when compensations must be reasoned about together.
- Saga choreography lets each service react to events: use it for short flows between 2-3 services that already publish.
- Every saga step needs its compensation defined before implementation. A step with no compensation is a step that must move to the end or become idempotent-retryable.
- Correlate everything: one id threads the whole flow through logs, events, and compensations.
Evolution
- Strangler fig: route one slice at a time behind a facade, run old and new in parallel, retire the old path only after traffic and data are verified.
- Branch by abstraction inside a service: introduce the seam, implement behind it, switch, delete the old side.
- Schema and event changes follow expand-contract, as in
system-design-data-architecture.
Anti-Patterns
- No CQRS by default: separate models only after one model demonstrably fails both sides.
- No event sourcing for ordinary CRUD: audit needs an audit log, not a rebuilt world.
- No dual write: outbox or CDC, never a database write followed by a hopeful publish.
- No choreography past 4 steps: nobody can answer "where is order 123" without a coordinator.
- No two-phase commit on the hot path: the coordinator is a SPOF and the locks are the outage.
- No distributed lock as a correctness mechanism: leases expire under GC pauses; use fencing tokens or single ownership.
References
- Integration Patterns Detail - outbox mechanics, saga compensation table, event versioning, BFF ownership, strangler phases