System Design Documentation
System design on one page — the interview method, key numbers, scaling, caching, databases, CAP, queues, reliability and question templates.
System design on one page: the method, the numbers, every building block with its trade-offs, and templates for common questions. New to this? Read System Design for Freshers first.
1. The Interview Method
- Requirements: functional (what it does) and non-functional (scale, latency, availability).
- Estimates: users, requests per second, storage.
- API: the main endpoints.
- Data model: tables and keys; SQL or NoSQL.
- High-level design: client → load balancer → servers → cache → database.
- Deep dive: one or two hard parts.
- Trade-offs: bottlenecks, failures, monitoring.
2. Numbers to Remember
| Quantity | Value |
|---|---|
| Seconds in a day | 86,400 ≈ 10⁵ |
| 1 million requests a day | ≈ 12 per second |
| Peak traffic | About 2–3× the average |
| RAM read | ~100 ns |
| SSD read | ~100 µs |
| Same data centre round trip | ~0.5 ms |
| India ↔ US round trip | ~200 ms |
| Availability | Downtime a year |
|---|---|
| 99% | 3.65 days |
| 99.9% | 8.76 hours |
| 99.99% | 52.6 minutes |
Estimation example: 10 M daily users × 20 reads = 200 M reads a day ≈ 2,000 per second, so about 5,000 at peak.
3. Scaling
| Vertical | Horizontal | |
|---|---|---|
| How | Bigger machine | More machines |
| Limit | Hardware | Almost none |
| Failure | Single point | Others keep serving |
| Needs | Nothing | Stateless servers, load balancer |
Stateless servers: keep sessions in Redis or JWTs and files in object storage, so any server can take any request. See System Design Basics: Scalability.
4. Load Balancing
| Algorithm | How |
|---|---|
| Round robin | Takes turns |
| Least connections | To the least busy server |
| IP hash | The same client goes to the same server |
| Weighted | More traffic to stronger servers |
L4 balances by IP and port (fast). L7 balances by URL, cookies and headers (smarter). Health checks remove dead servers. See Load Balancing.
5. Caching
| Pattern | How |
|---|---|
| Cache-aside | App checks cache; on a miss reads DB and fills cache (most common) |
| Write-through | Writes go to cache and DB together |
| Write-back | Write to cache, DB later (fast, risky) |
Eviction: LRU, LFU, TTL. Problems: stale data, stampede on a hot key, penetration by missing keys. CDN caches static files near users. See Caching, CDN and Databases at Scale.
6. Databases
| Need | Choose |
|---|---|
| Transactions, joins | SQL (PostgreSQL, MySQL) |
| Flexible documents | MongoDB |
| Very high write volume | Cassandra |
| Cache, counters | Redis |
| Relationships | Neo4j |
| Search | Elasticsearch |
| Files, images | S3 (never in the DB) |
| Technique | Solves |
|---|---|
| Index | Slow reads |
| Read replicas | Too many reads |
| Sharding | Too many writes, or too much data |
| Denormalisation | Expensive joins |
| Connection pooling | Too many connections |
Sharding: range (hot spots), hash (hard to add shards), consistent hashing (moves few keys). See Sharding and Replication.
7. CAP and Consistency
During a network partition, choose consistency or availability.
| Choice | Behaviour | Good for |
|---|---|---|
| CP | Refuse rather than return stale data | Payments, inventory |
| AP | Always answer, maybe stale | Feeds, likes |
Strong consistency: reads see the latest write. Eventual consistency: replicas catch up soon. See CAP Theorem.
8. Asynchronous Work
A message queue (Kafka, RabbitMQ, SQS) separates producers from consumers. Use it for emails, video processing, and absorbing traffic spikes.
| Queue (RabbitMQ) | Log (Kafka) | |
|---|---|---|
| After reading | Message removed | Message kept, can replay |
| Good for | Task jobs | Event streams |
Delivery is usually at-least-once, so consumers must be idempotent. See Message Queues.
9. Communication
| Option | Use |
|---|---|
| REST | Most APIs |
| GraphQL | Clients choosing their fields |
| gRPC | Fast service-to-service |
| WebSocket | Two-way real time (chat) |
| SSE | Server-to-client stream |
| Polling | Simple, slightly delayed updates |
10. Reliability and Protection
| Pattern | Purpose |
|---|---|
| Redundancy | Remove single points of failure |
| Replication + failover | Survive a server dying |
| Rate limiting | Stop abuse (token bucket, sliding window) |
| Circuit breaker | Stop calling a failing service |
| Retry with backoff | Recover from brief failures |
| Timeouts | Don't wait forever |
| Idempotency keys | Make retries safe (payments) |
| Monitoring | Metrics, logs, traces, alerts |
11. Monolith vs Microservices
| Monolith | Microservices | |
|---|---|---|
| Deploy | One unit | Each service alone |
| Start | Simple | Complex |
| Scale | Everything together | Per service |
| Needs | Little | Gateway, discovery, tracing |
Start with a monolith and split only when teams or scale demand it. API gateways handle auth, routing and rate limits. Saga handles transactions across services.
12. Question Templates
| Question | Key ideas |
|---|---|
| URL shortener | Base62 IDs, cache, read-heavy |
| Chat app | WebSockets, message store, presence |
| News feed | Fan-out on write vs read, cache |
| Rate limiter | Token bucket in Redis |
| Video platform | Object storage, transcoding queue, CDN |
| Notification system | Queue, workers, retries |
| Ride sharing | Geo index, matching, live location |
| File storage | Chunks, metadata DB, S3 |
Practice: explain your own project's architecture and how it would scale to 1 lakh users.
13. Interview Questions
How do you scale a read-heavy system? Cache, read replicas and a CDN.
How do you scale writes? Sharding, queues and batching.
What is a single point of failure? A part whose failure stops everything; remove it with redundancy.
What is consistent hashing? Mapping keys and servers on a ring, so changes move few keys.
Latency vs throughput? Time per request vs requests per second.
SQL or NoSQL? SQL for relations and transactions; NoSQL for flexible schemas and huge scale.
Why use a message queue? To decouple services, smooth spikes and run slow work later.
What is idempotency? Repeating an action has the same effect as doing it once.
14. Cheat Sheet & References
Method
Requirements → Estimates → API → Data → High-level → Deep dive → Trade-offs
Quick maths
1 day ≈ 10^5 s · 1M/day ≈ 12/s · peak ≈ 3× average
99.9% ≈ 8.8 h down a year · 99.99% ≈ 53 min
Fixes
Slow reads → index, cache, replicas, CDN
Slow writes → sharding, queue, batching
Spikes → queue, autoscaling, rate limit
Failures → redundancy, retries, circuit breaker
Standard architecture
Client → DNS → CDN → Load balancer → App servers → Cache → DB
↓
Queue → Workers