design-microservices

Design Microservices

Decompose a system into independently deployable services aligned to business capabilities using bounded contexts.

Why This Is Best Practice

Adopted by: Netflix, Amazon, Uber, Spotify — all publicly documented their migrations from monoliths Impact: Amazon's service decomposition enabled teams to deploy independently, reducing release cycle from months to minutes; Netflix processes 2+ billion API requests/day across ~700 microservices.

Why best: DDD bounded contexts provide the natural seam for service decomposition — they align services to business domains, minimize cross-service coupling, and give each team a clear ownership boundary. Sam Newman's patterns (strangler fig, anti-corruption layer) provide proven migration paths.

Steps

1. Map business capabilities — Identify the core domains (e.g., Orders, Inventory, Payments) using event storming or domain mapping workshops.
2. Define bounded contexts — Draw context boundaries where the ubiquitous language changes; each context is a candidate service.
3. Apply the right decomposition pattern — Decompose by business capability (preferred) or subdomain; avoid decomposing by technical layer (e.g., "database service").
4. Define service contracts — Specify each service's public API (REST, gRPC, or events) before implementation; treat contracts as stable interfaces.
5. Design for failure — Every inter-service call can fail; apply timeouts, retries with backoff, and circuit breakers (Hystrix/Resilience4j pattern).
6. Choose a data strategy — Each service owns its data store; share data via events or APIs, never via shared databases.
7. Plan observability — Agree on distributed tracing (OpenTelemetry), structured logging, and health endpoints before first service ships.

Rules

• One database per service — shared databases couple services and eliminate independent deployability.
• Services must be independently deployable without coordinating with other teams.
• Avoid nano-services — if a service only exists to wrap a CRUD table, it belongs in a larger service.
• Prefer choreography (events) over orchestration for business workflows to avoid a central point of failure.

Examples

E-commerce decomposition:

• order-service — creates and tracks orders; publishes OrderPlaced events.
• inventory-service — listens for OrderPlaced, reserves stock, publishes StockReserved.
• payment-service — listens for StockReserved, charges customer.

Each service has its own DB; communication is via Kafka events; no service calls another's DB directly.

Common Mistakes

• Distributed monolith — services that must deploy together or share a database gain none of the benefits of microservices.
• Decomposing too early — start with a well-structured monolith ("modular monolith") until domain boundaries are understood.
• No contract versioning — breaking API changes cascade across services; version APIs from day one.

When NOT to Use

• When the team has fewer than 8 engineers, the operational overhead of independent deployments, distributed tracing, and per-service databases will consume more capacity than the autonomy gains.
• When the domain boundaries are not yet understood — building microservices before the domain model stabilizes results in services that must be merged or split within months, producing a distributed monolith with extra steps.
• When the system has no requirement for independent scaling or independent team ownership, a well-structured modular monolith delivers the same maintainability at a fraction of the operational complexity.