12 Design Patterns for Building Scalable Systems#

These are the patterns you see in every large-scale system — from Netflix to Stripe to Uber. Each solves a specific scaling challenge. Learn when to use each and when to avoid it.

1. CQRS (Command Query Responsibility Segregation)#

Separate read and write models:

Writes: API → Command Handler → Write DB (normalized)
                                    ↓ events
Reads:  API → Query Handler → Read DB (denormalized views)

When: Reads and writes have very different patterns (e.g., 1000:1 read-to-write ratio). Each can scale independently.

Example: E-commerce product catalog. Writes are rare (admin updates), reads are massive (millions of shoppers).

2. Event Sourcing#

Store events, not state:

Event Store:
  OrderCreated { items: [...], total: 100 }
  PaymentProcessed { amount: 100 }
  ItemShipped { tracking: "XY123" }

Current state = replay all events

When: You need a complete audit trail, time-travel debugging, or the ability to rebuild state from scratch.

Example: Financial systems, order processing, compliance-heavy domains.

3. Saga Pattern#

Manage distributed transactions without 2PC:

Choreography:
  Order → Payment → Inventory → Shipping
  (each service reacts to the previous event)

Orchestration:
  Saga Coordinator tells each service what to do
  On failure → compensating actions (refund, unreserve)

When: Transactions span multiple services and you need eventual consistency with rollback capability.

4. Circuit Breaker#

Fail fast instead of waiting for broken services:

Closed → errors exceed threshold → Open (fail fast for 30s)
Open → timer expires → Half-Open (try one request)
Half-Open → success → Closed | failure → Open

When: Calling external services or downstream microservices that may be unreliable.

5. Bulkhead#

Isolate failures by compartmentalizing resources:

Thread Pool A (20) → Payment Service
Thread Pool B (10) → Email Service (if slow, only Pool B affected)
Thread Pool C (5)  → Analytics (non-critical)

When: Different services have different criticality levels. A slow analytics call shouldn't block payments.

6. Sidecar Pattern#

Deploy helper functionality alongside your service:

┌─────────────────────┐
│ Your Service         │
│ (business logic)     │
│                      │
│ ┌──────────────────┐ │
│ │ Sidecar (Envoy)  │ │
│ │ - mTLS           │ │
│ │ - Logging        │ │
│ │ - Metrics        │ │
│ │ - Retries        │ │
│ └──────────────────┘ │
└─────────────────────┘

When: You want cross-cutting concerns (security, observability) without changing application code. Foundation of service mesh (Istio, Linkerd).

7. Strangler Fig#

Incrementally migrate from monolith to microservices:

Phase 1: All traffic → Monolith
Phase 2: /users → New User Service | rest → Monolith
Phase 3: /orders → New Order Service | /users → User Service | rest → Monolith
Phase N: Monolith is empty → delete it

When: Migrating a legacy system. Never rewrite from scratch — strangle it gradually.

8. API Gateway / BFF#

Single entry point for all clients:

Mobile → Mobile BFF → aggregates 3 services → optimized response
Web    → Web BFF    → aggregates 5 services → full response

When: Multiple clients need different views of the same data. Reduces client-side complexity and network calls.

9. Database per Service#

Each microservice owns its data:

User Service → Users DB (PostgreSQL)
Order Service → Orders DB (PostgreSQL)
Search Service → Search Index (Elasticsearch)
Cache Service → Cache (Redis)

When: Services need independent scaling and deployment. No shared database = no coupling.

Trade-off: Cross-service queries require API calls or events, not JOINs.

10. Outbox Pattern#

Reliably publish events when database changes:

Transaction:
  1. INSERT INTO orders (id, ...) VALUES (...)
  2. INSERT INTO outbox (event_type, payload) VALUES ('OrderCreated', {...})

Background worker:
  Poll outbox → publish to Kafka → mark as published

When: You need to update a database AND publish an event atomically. Prevents the dual-write problem.

11. Materialized View#

Pre-compute query results for fast reads:

Write: Order created → event → Materializer
  → Updates "dashboard_summary" table:
    total_orders: 1542
    revenue_today: $45,230
    avg_order_value: $29.33

Read: SELECT * FROM dashboard_summary (1ms, no JOINs)

When: Complex queries are too slow to run in real-time. Pre-compute and serve from a denormalized view.

12. Backpressure#

Slow down producers when consumers can't keep up:

Without backpressure:
  Producer (1000/sec) → Queue (growing!) → Consumer (100/sec) → OOM crash

With backpressure:
  Producer → Queue (near capacity) → signal producer to slow down
  Producer (100/sec) → Queue (stable) → Consumer (100/sec) ✓

When: Producers can generate data faster than consumers can process. Essential for streaming systems.

Pattern Decision Matrix#

Architecture: Putting It Together#

Scalable E-Commerce Platform#

Mobile App → API Gateway (BFF pattern)
Web App   → API Gateway

API Gateway → Circuit Breaker → Product Service (CQRS)
                                  → Write Model (PostgreSQL)
                                  → Read Model (Elasticsearch)
           → Circuit Breaker → Order Service (Saga + Event Sourcing)
                                  → Event Store
                                  → Outbox → Kafka
           → Circuit Breaker → Payment Service (Bulkhead isolated)

Kafka → Materialized View updater → Dashboard DB
     → Analytics pipeline → ClickHouse

Service Mesh (Sidecar): mTLS, retries, observability

Generate this architecture →

Anti-Patterns#

1. Using all 12 patterns at once — start simple, add patterns when you feel pain
2. Event sourcing everywhere — most services don't need it (CRUD is fine)
3. Saga without compensation — if you can't rollback, don't use a saga
4. CQRS for simple CRUD — adds complexity with no benefit for simple apps
5. Premature microservices — start monolith, extract when teams grow

Summary#

Start with the simplest architecture that works. Add patterns when you hit specific scaling problems:

• First scaling pain: Add caching + read replicas
• Service independence: Database per service + API gateway
• Reliability: Circuit breaker + bulkhead + retries
• Async processing: Event-driven + outbox + Kafka
• Complex domains: CQRS + event sourcing + sagas
• Migration: Strangler fig (never rewrite)

---

Design scalable architectures at codelit.io — 100 product specs, 90+ templates, 7 audit tools, 29 export formats. From specs to production.