Microservices Architecture

Patterns, trade-offs, and practical guidance for designing, building, and operating microservices systems.

---

When to Use Microservices (and When Not To)

Microservices are a good fit when:

• Multiple teams need to deploy independently on different release cadences
• Different parts of the system have fundamentally different scaling requirements
• The domain is well understood and service boundaries are clear
• The organization can invest in infrastructure automation, monitoring, and CI/CD

A monolith is better when:

• The team is small (fewer than 8-10 engineers)
• The domain is not yet well understood and boundaries are likely to shift
• Speed of initial development matters more than independent deployability
• The organization lacks mature DevOps practices
• The application has low complexity and uniform scaling needs

Start with a well-structured monolith. Extract services only when you have a clear, justified reason. Premature decomposition creates distributed monoliths.

---

Service Boundaries

Domain-Driven Design (DDD)

Service boundaries should align with business domains, not technical layers.

Bounded Contexts define the boundaries within which a particular domain model applies. Each microservice should map to one bounded context.

Order Context          Inventory Context        Shipping Context
+-----------------+    +-----------------+      +-----------------+
| Order           |    | Product         |      | Shipment        |
| OrderLine       |    | StockLevel      |      | Carrier         |
| Payment         |    | Warehouse       |      | TrackingEvent   |
+-----------------+    +-----------------+      +-----------------+

Guidelines for drawing boundaries:

• Business capability: Each service represents a business function (ordering, billing, shipping)
• Data ownership: A service owns its data and exposes it only through its API
• Autonomy: A service can be developed, deployed, and scaled independently
• Cohesion: Related behavior lives together; unrelated behavior lives apart
• Loose coupling: Changes in one service should rarely require changes in another

Context Mapping

Define explicit relationships between bounded contexts:

• Shared Kernel: Two contexts share a small common model (use sparingly)
• Customer-Supplier: Upstream context provides what downstream needs
• Anticorruption Layer: Translates between contexts to prevent model leakage
• Conformist: Downstream adopts upstream model as-is

---

Communication Patterns

Synchronous Communication

REST (HTTP/JSON)

• Widely understood, good tooling, human-readable
• Suitable for CRUD operations and simple request-response flows
• Versioning via URL path (/v1/orders) or headers
• Use when: clients need immediate responses, operations are simple

gRPC (Protocol Buffers)

• Binary protocol, significantly faster than JSON over HTTP
• Strongly typed contracts via .proto files
• Supports streaming (unary, server-streaming, client-streaming, bidirectional)
• Use when: low-latency internal service-to-service communication, high throughput

service OrderService {
  rpc CreateOrder (CreateOrderRequest) returns (OrderResponse);
  rpc StreamOrderUpdates (OrderQuery) returns (stream OrderUpdate);
}

Asynchronous Communication

Message Queues (Point-to-Point)

• One producer sends a message; one consumer processes it
• Work distribution, task offloading, load leveling
• Examples: RabbitMQ queues, SQS

Events (Publish-Subscribe)

• One producer publishes an event; multiple consumers react independently
• Decouples producers from consumers
• Examples: Kafka topics, SNS, RabbitMQ exchanges

Choose async when:

• The caller does not need an immediate result
• You want to decouple services temporally
• You need to buffer load spikes
• Multiple consumers need to react to the same event

---

API Gateway Pattern

Mobile App  --\
Web App     ----> API Gateway ----> Order Service
Partner API --/        |   \------> Inventory Service
                       |   \------> User Service
                       v
                Authentication
                Rate Limiting
                Request Routing
                Response Aggregation
                Protocol Translation
                Logging / Metrics

Responsibilities:

• Routing: Direct requests to the correct backend service
• Authentication/Authorization: Validate tokens, enforce policies at the edge
• Rate limiting: Protect backend services from traffic spikes
• Response aggregation: Combine data from multiple services into a single response
• Protocol translation: Accept REST from clients, forward as gRPC internally
• SSL termination: Handle TLS at the gateway

Implementations:

• Kong, NGINX, AWS API Gateway, Envoy, Traefik

Backend-for-Frontend (BFF):

A variant where each client type (mobile, web, partner) gets its own gateway tailored to its needs, avoiding a one-size-fits-all gateway.

---

Service Discovery

Services need to locate each other at runtime when instances scale dynamically.

Client-Side Discovery

The client queries a service registry and selects an instance.

Client -> Service Registry -> [Instance A, Instance B, Instance C]
Client -> Instance B (selected via load balancing)

Server-Side Discovery

The client sends requests to a load balancer, which queries the registry.

Client -> Load Balancer -> Service Registry -> Instance A

Implementations:

• Consul: Service registry with health checking, key-value store, DNS interface
• Eureka: Netflix OSS, common in Spring Cloud ecosystems
• Kubernetes DNS: Built-in service discovery via DNS names (e.g., order-service.default.svc.cluster.local)
• AWS Cloud Map: Managed service discovery for AWS workloads

In Kubernetes environments, built-in DNS-based discovery often eliminates the need for a separate registry.

---

Event-Driven Architecture

Event Sourcing

Instead of storing current state, store the sequence of events that led to the current state.

Events (append-only log):
1. OrderCreated { orderId: 42, items: [...], total: 150.00 }
2. PaymentReceived { orderId: 42, amount: 150.00 }
3. OrderShipped { orderId: 42, trackingId: "XYZ123" }

Current state is derived by replaying events.

Benefits: Full audit trail, temporal queries ("what was the state at time T"), ability to rebuild read models, supports event replay for debugging.

Costs: Increased storage, more complex queries, eventual consistency, schema evolution for events requires care.

CQRS (Command Query Responsibility Segregation)

Separate the write model (commands) from the read model (queries).

Commands (writes)            Queries (reads)
     |                            |
     v                            v
Write Model  --events-->   Read Model (projection)
(normalized)               (denormalized, optimized for queries)

• The write side validates and persists domain events
• The read side subscribes to events and maintains query-optimized views
• Often paired with event sourcing, but they are independent patterns
• Use when read and write workloads have very different performance or modeling needs

---

Message Brokers

When to use which:

Broker	Best for	Key characteristics
RabbitMQ	Task queues, work distribution, routing	AMQP protocol, flexible routing (direct, topic, fanout), message acknowledgment, per-message delivery guarantees, lower throughput than Kafka
Apache Kafka	Event streaming, event sourcing, high-throughput pipelines	Append-only log, consumer groups, message retention by time/size, replay capability, horizontal scaling via partitions, high throughput
Amazon SQS	Simple cloud-native queuing	Fully managed, no infrastructure to operate, standard and FIFO queues, dead-letter queues, integrates with AWS Lambda

Decision guidance:

• Need message replay or event sourcing? Kafka
• Need complex routing logic (topic exchanges, headers-based routing)? RabbitMQ
• Want zero infrastructure management on AWS? SQS
• Need strict ordering with exactly-once delivery? SQS FIFO or Kafka (with idempotent producers)
• Need pub/sub with persistent consumer groups? Kafka
• Need lightweight, traditional work queues? RabbitMQ or SQS

---

Data Management

Database per Service

Each service owns its database. No other service accesses it directly.

Order Service  --> Orders DB (PostgreSQL)
Product Service --> Products DB (MongoDB)
Search Service --> Search Index (Elasticsearch)

Benefits: Independent schema evolution, technology choice per service, no shared-database coupling.

Challenge: Cross-service queries and distributed transactions.

Saga Pattern

Manage distributed transactions as a sequence of local transactions with compensating actions.

Choreography (event-based):

Order Service: CreateOrder --> emit OrderCreated
Payment Service: hears OrderCreated --> charge payment --> emit PaymentCompleted
Inventory Service: hears PaymentCompleted --> reserve stock --> emit StockReserved
Order Service: hears StockReserved --> confirm order

On failure at any step: each service listens for failure events and runs compensation
(e.g., PaymentFailed --> refund payment, release stock)

Orchestration (coordinator-based):

Saga Orchestrator:
  1. Tell Order Service: create order
  2. Tell Payment Service: charge payment
  3. Tell Inventory Service: reserve stock
  On failure: send compensating commands in reverse order

• Choreography: simpler for few steps, harder to trace as complexity grows
• Orchestration: explicit flow control, easier to reason about, single point to monitor

Eventual Consistency

In a microservices system, strong consistency across services is impractical. Accept that:

• Data will be temporarily inconsistent between services
• Design UIs to handle in-progress states gracefully
• Use idempotent operations to safely retry
• Implement read-your-writes consistency where user experience demands it

---

Circuit Breaker Pattern

Prevent cascading failures when a downstream service is unhealthy.

States:

CLOSED (normal) --failures exceed threshold--> OPEN (failing fast)
OPEN --timeout expires--> HALF-OPEN (testing)
HALF-OPEN --success--> CLOSED
HALF-OPEN --failure--> OPEN

Implementation concerns:

Retry with exponential backoff:

import time

def call_with_retry(func, max_retries=3, base_delay=1.0):
    for attempt in range(max_retries):
        try:
            return func()
        except TransientError:
            if attempt == max_retries - 1:
                raise
            delay = base_delay * (2 ** attempt)  # 1s, 2s, 4s
            time.sleep(delay)

Timeout: Set timeouts on every external call. A missing timeout is a latency leak. Calculate cascading timeouts so upstream timeouts are longer than downstream.

Fallback strategies:

• Return cached data (stale but available)
• Return a default/degraded response
• Queue the request for later processing
• Return an error with clear guidance

Libraries: Resilience4j (Java), Polly (.NET), Hystrix (legacy), custom implementations.

---

Distributed Tracing and Observability

Three Pillars of Observability

Logs: Structured, machine-parsable records of discrete events.

• Use structured logging (JSON) with consistent fields
• Include a correlation/trace ID in every log entry
• Centralize with ELK (Elasticsearch, Logstash, Kibana), Loki, or Splunk

Metrics: Numeric measurements aggregated over time.

• RED method: Rate, Errors, Duration (for request-driven services)
• USE method: Utilization, Saturation, Errors (for resources)
• Tools: Prometheus + Grafana, Datadog, CloudWatch

Traces: End-to-end request paths across services.

• Propagate trace context (W3C Trace Context or B3 headers) across service boundaries
• Each service creates spans representing its work
• Tools: Jaeger, Zipkin, AWS X-Ray, OpenTelemetry (vendor-neutral standard)

OpenTelemetry

The emerging standard for instrumentation. Provides APIs and SDKs for traces, metrics, and logs with exporters to multiple backends. Prefer OpenTelemetry over vendor-specific instrumentation.

Health checks

Every service should expose:

• Liveness: "Is the process running?" (restart if not)
• Readiness: "Can the service handle traffic?" (remove from load balancer if not)

---

Service Mesh

A dedicated infrastructure layer for managing service-to-service communication.

What it solves:

• Mutual TLS between services without application code changes
• Traffic management (retries, timeouts, circuit breaking) at the infrastructure level
• Observability (metrics, traces) injected automatically via sidecar proxies
• Traffic splitting for canary deployments
• Access control policies between services

How it works:

Service A -> Sidecar Proxy A ---network---> Sidecar Proxy B -> Service B
                    \                              /
                     -----> Control Plane <--------
                            (configuration, certificates, policies)

Implementations:

• Istio: Feature-rich, Envoy-based, higher operational complexity
• Linkerd: Lightweight, simpler to operate, Rust-based proxy
• AWS App Mesh: Managed mesh for AWS workloads

When to adopt:

• You have many services (dozens+) and need uniform observability and security
• You want to enforce mTLS without changing application code
• You need advanced traffic management (fault injection, mirroring)
• Do not adopt for fewer than 5-10 services; the overhead is not justified

---

Deployment Patterns

Blue-Green Deployment

• Maintain two identical production environments (blue and green)
• Deploy new version to the idle environment
• Switch traffic from blue to green (or vice versa) once verified
• Instant rollback by switching traffic back
• Cost: maintaining two full environments

Canary Deployment

• Route a small percentage of traffic (e.g., 5%) to the new version
• Monitor error rates, latency, and business metrics
• Gradually increase traffic if metrics are healthy
• Roll back by routing all traffic to the old version
• Lower risk than blue-green, but slower rollout

Rolling Deployment

• Replace instances of the old version one at a time (or in batches)
• At any point, both old and new versions are running simultaneously
• No additional infrastructure cost
• Requires backward-compatible changes (APIs, database schemas)
• Slower rollback compared to blue-green

Choosing a strategy:

• Need instant rollback? Blue-green
• Want to minimize risk with gradual exposure? Canary
• Want simplicity and low infrastructure cost? Rolling

---

Container Orchestration Basics

Microservices are typically deployed as containers managed by an orchestrator.

Kubernetes core concepts:

• Pod: Smallest deployable unit; one or more containers sharing network and storage
• Deployment: Declares desired state (image version, replica count); handles rolling updates
• Service: Stable network endpoint that routes to healthy pods
• ConfigMap / Secret: Externalized configuration and sensitive data
• Horizontal Pod Autoscaler (HPA): Scales pod count based on CPU, memory, or custom metrics
• Namespace: Logical isolation within a cluster

Alternatives:

• AWS ECS/Fargate: Managed container orchestration without Kubernetes complexity
• Docker Swarm: Simpler orchestration, suitable for smaller deployments
• Nomad: HashiCorp orchestrator supporting containers and non-container workloads

---

Testing Microservices

Testing pyramid for microservices:

           /  E2E Tests  \        (few, slow, expensive)
          / Integration    \
         / Contract Tests    \
        / Component Tests      \
       /   Unit Tests            \   (many, fast, cheap)

Contract Testing

Verify that service interfaces remain compatible without running full integration tests.

• Consumer-driven contracts: The consumer defines expectations; the provider verifies it meets them
• Tools: Pact, Spring Cloud Contract
• Catches breaking API changes early in CI

Component Testing

Test a single service in isolation with its dependencies stubbed or mocked.

Integration Testing

Test interactions between real services in a shared test environment.

• Expensive to maintain; use sparingly
• Focus on critical paths and failure modes

End-to-End Testing

Test complete user journeys across the full system.

• Slowest and most brittle; keep the count low
• Supplement with synthetic monitoring in production

Testing in production:

• Feature flags to control rollout
• Canary deployments with automated metric checks
• Synthetic transactions (health-check requests that exercise real paths)
• Chaos engineering (deliberately inject failures to verify resilience)

---

Common Anti-Patterns

Distributed Monolith

Services are deployed independently but must be changed, tested, and deployed together. Causes:

• Shared libraries with business logic
• Synchronous call chains that create tight runtime coupling
• Shared database schemas

Shared Database

Multiple services read/write the same database tables. Eliminates independent deployability and schema evolution. Always give each service its own data store.

Chatty Services

Excessive fine-grained calls between services. If Service A makes 20 calls to Service B to fulfill one request:

• Aggregate into coarser-grained APIs
• Consider if the boundary is drawn incorrectly
• Use batch endpoints or data replication

Mega Service / God Service

One service grows to handle too many responsibilities, becoming a monolith in disguise. Enforce bounded contexts.

Ignoring Network Fallibility

Treating remote calls like local function calls. Always account for latency, partial failure, and network partitions.

Synchronous Chains

A -> B -> C -> D where each call is synchronous. Latency compounds, and failure in any service fails the entire chain. Prefer async communication or reduce chain depth.

---

Migration from Monolith

Strangler Fig Pattern

Incrementally replace monolith functionality with microservices.

Phase 1: All traffic --> Monolith

Phase 2: Proxy/Router --> New Service (handles /orders)
                     \--> Monolith (handles everything else)

Phase 3: Proxy/Router --> Order Service
                     \--> Payment Service
                     \--> Monolith (shrinking)

Phase N: Monolith is decommissioned

Steps:

1. Place a routing layer (API gateway, reverse proxy) in front of the monolith
2. Identify a bounded context to extract
3. Build the new service alongside the monolith
4. Redirect traffic for that context to the new service
5. Repeat, shrinking the monolith over time

Branch by Abstraction

Refactor internally before extracting.

1. Create an abstraction (interface) for the functionality to extract
2. Implement the abstraction with the existing monolith code
3. Build a second implementation that calls the new microservice
4. Switch the implementation (feature flag or configuration)
5. Remove the old implementation once the new service is stable

Migration guidance:

• Extract the most independently valuable bounded context first
• Avoid extracting shared/core logic early (it has the most coupling)
• Ensure the monolith and new services can coexist indefinitely
• Data migration is the hardest part; plan for dual-write or change-data-capture strategies
• Measure success by deployment independence, not service count

---

References

• Sam Newman, Building Microservices (2nd edition)
• Chris Richardson, Microservices Patterns
• Martin Fowler, microservices articles at martinfowler.com
• Eric Evans, Domain-Driven Design
• OpenTelemetry documentation: opentelemetry.io
• Kubernetes documentation: kubernetes.io