nwave-ai/nw-sd-framework

System Design Framework

The 4-Step Process

Every system design follows this structure. Skipping steps is the top mistake.

Step 1: Understand the Problem and Establish Design Scope (3-10 min)

Narrow an impossibly broad question into a tractable problem.

Ask about: users and scale | most important features | read/write ratio | non-functional requirements (latency, availability, consistency) | existing infrastructure | special constraints (mobile-first, offline, regulatory)

Produce: functional requirements (3-5 bullets) | non-functional requirements (scale, latency, availability, consistency model) | capacity estimation (QPS, storage, bandwidth)

Red flags if skipped: designing a system nobody asked for | over-engineering for imaginary scale | missing critical constraints (GDPR, real-time)

Step 2: Propose High-Level Design and Get Buy-In (10-15 min)

Sketch the big picture. Validate before diving deep.

Do: draw architecture diagram (clients, servers, databases, caches, queues) | define API contract (REST/GraphQL/gRPC -- key endpoints) | design data model (entities, relationships, access patterns) | walk through 1-2 core use cases end-to-end | get buy-in: "Does this make sense before I go deeper?"

API patterns: RESTful for CRUD-heavy | GraphQL for flexible client queries | gRPC for internal service-to-service | WebSocket/SSE for real-time

Data model: SQL vs NoSQL based on access patterns, not hype | denormalization trade-offs | partitioning key selection (directly impacts scalability)

Step 3: Design Deep Dive (10-25 min)

Go deep on 2-3 components.

Choose: most technically challenging | most interesting trade-offs | bottleneck components (highest load, most failure-prone)

Depth means: specific algorithms (consistent hashing, Bloom filters) | failure modes and handling | scaling strategy per component | data flow with edge cases | monitoring and operational concerns

Step 4: Wrap Up (3-5 min)

Cover: summarize design in 2-3 sentences | identify known bottlenecks | what you'd improve with more time | operational concerns (monitoring, alerting, deployment) | future enhancements

Avoid: introducing entirely new components at this stage | second-guessing your design

---

Back-of-Envelope Estimation

Powers of 2 Reference

| Power | Value | Meaning | |-------|-------|---------| | 10 | 1 Thousand | 1 KB | | 20 | 1 Million | 1 MB | | 30 | 1 Billion | 1 GB | | 40 | 1 Trillion | 1 TB | | 50 | 1 Quadrillion | 1 PB |

Latency Numbers Every Engineer Should Know

| Operation | Latency | |-----------|---------| | L1 cache reference | 0.5 ns | | L2 cache reference | 7 ns | | Main memory reference | 100 ns | | Compress 1KB (Zippy) | 10 us | | Send 2KB over 1 Gbps | 20 us | | Read 1 MB from memory | 250 us | | Datacenter round trip | 500 us | | Disk seek | 10 ms | | Read 1 MB from network | 10 ms | | Read 1 MB from disk | 30 ms | | CA to Netherlands round trip | 150 ms |

Key takeaways: memory fast, disk slow -- cache aggressively | compress before network send | inter-datacenter trips expensive -- minimize cross-region calls

Common Estimation Patterns

DAU to QPS: QPS = DAU * actions_per_user / 86400 | Peak QPS = QPS * 2 (or *3 for spiky)

Storage: daily = DAU * actions * avg_size | yearly = daily * 365 | 5-year = yearly * 5

Bandwidth: QPS * average_response_size

Servers: Peak QPS / QPS_per_server where CPU-bound ~hundreds | IO-bound with cache ~thousands | static content ~tens of thousands

Estimation Example: Twitter-like Service

150M DAU, 2 tweets/day, 10 reads/day
Write QPS = 150M * 2 / 86400 ~ 3,500
Read QPS = 150M * 10 / 86400 ~ 17,000; Peak ~ 50,000
Storage: 300M tweets * 1KB + 30M media * 500KB ~ 15.3 TB/day

---

Scaling Ladder

Each step solves a specific bottleneck. Never introduce a component without articulating which bottleneck it addresses.

1. Load balancer -- distribute traffic across web servers
2. Database replication -- master-slave for read scaling
3. Cache layer -- reduce database load (Redis/Memcached)
4. CDN -- serve static content from edge
5. Stateless web tier -- move session state to shared store
6. Database sharding -- horizontal partitioning for write scaling
7. Message queue -- decouple components, handle spikes
8. Logging, metrics, monitoring -- observability at scale
9. Multiple data centers -- geographic redundancy and latency reduction

---

Common Pitfalls

1. Jumping to solutions -- design before understanding requirements
2. Over-engineering -- adding components for imaginary scale
3. Ignoring trade-offs -- every choice has a cost; name it
4. SPOF blindness -- always ask "what if this dies?"
5. Neglecting data -- the data model drives everything
6. Forgetting operations -- a system you can't monitor is one you can't run
7. Not doing math -- gut feelings are wrong; estimates keep you honest