Distributed systems revision

Revision guide · ~25 min · dense bullets, not full lessons

Revision for Distributed Systems & Scaling.

Load balancing

• Spread traffic: L4 (TCP) vs L7 (HTTP, path/host). Health checks + connection draining.
• Algorithms: round-robin, least-conn, consistent hashing (sticky-ish without sticky hell).
• LB is not magic HA — backends + data layer still fail.

Horizontal vs vertical

• Vertical: bigger box — simple ceiling. Horizontal: more boxes — needs statelessness or shared state design.
• Scale the bottleneck you measured, not the one you fear.

Monolith vs microservices

• Monolith: one deploy, simple transactions, easy refactor inside process.
• Microservices: independent deploy/scale; pay with network, dual writes, distributed tracing, org overhead.
• Start modular monolith; extract when a team/boundary/scalability pressure is real.

Service communication

• Sync HTTP/gRPC: simple request path; cascading failures if unbounded.
• Async events/queues: temporal decoupling; eventual consistency; harder mental model.
• Timeouts, retries with jitter, bulkheads, circuit breakers — default when calling others.

Message queues

• Producer → broker → consumers. At-least-once delivery is common → idempotent consumers.
• Competing consumers scale throughput. Poison messages → DLQ.
• Ordering: per-partition / per-key, not global free lunch.
• Use for: emails, webhooks, fan-out, smoothing spikes — not as a database.

25-minute drill

1. Sketch request path: client → CDN → LB → 3 app nodes → primary+replica DB.
2. Argue monolith vs services for a 3-person team.
3. Why “exactly once” processing is usually “at-least-once + idempotency.”