System Design Workbook

A repeatable interview framework — clarify, estimate, contract, high-level, data, detail, evaluate. Read this first; every system writeup on this site follows it.

What an interviewer means when they say 'service', 'node', 'cluster' — the unit boundaries that shape every design.

Network abstractions over function-call semantics: gRPC, Thrift, REST-as-RPC, the leaks they hide and don't.

From strong to eventual: linearizability, sequential, causal, monotonic, eventual — what each costs and what each buys.

Crash-stop, crash-recovery, omission, Byzantine — the assumption stack under every fault-tolerance claim.

Latencies, throughputs, sizes, and conversion shortcuts to keep in your head for back-of-envelope math.

Four end-to-end estimation walk-throughs (Twitter, YouTube, WhatsApp, a search index) — the math anchored to the Foundations cheatsheet.

8-step, 4-step, SNAKE, PACELC-flavored — what each emphasizes and which to anchor on under pressure.

Nines as a budget, redundancy strategies, failover modes, and the cost of each 9.

Probability of working under expected conditions; how it differs from availability and where the trade-offs live.

Vertical, horizontal, and elasticity. Why naive scaling stalls and which axis to pick first.

The operational burden of every architectural decision: observability, deployability, on-call load.

Designing for the worst expected failure and degrading gracefully past it.

Which non-functional requirements to surface, in what order, and how to convert vague asks into numbers.

Hierarchical name resolution, caching, TTL trade-offs, and DNS as a load-balancing primitive.

L4 vs L7, global vs local, algorithms (round-robin, least-connections, consistent-hash), placement tiers.

Relational vs document vs wide-column vs graph: when each shape fits, and the trade-off triangle.

Consistent-hash ring, replication factor, versioning (vector clocks), failure detection. Dynamo-style design.

Edge caching, origin shielding, push vs pull, cache invalidation, signed URLs.

Globally-ordered unique IDs with causality. Snowflake, TrueTime, hybrid logical clocks.

Metrics, logs, traces — the three pillars and the data structures that scale each.

Real-time error capture, deduplication, alerting, blast-radius scoping.

Browser and mobile error capture, sampling, PII scrubbing, beacon transport.

Memcached vs Redis, sharding, eviction policies, replication, stampede protection.

FIFO vs at-least-once vs exactly-once, partitions, consumer groups, dead-letter queues.

Topic-based fan-out, ordering guarantees, filtering, retention, and the gap between pub-sub and queues.

Cap request rates per client to protect downstream services. Token bucket vs leaky bucket vs sliding window, with the gotchas of distributed coordination.

Object storage with metadata indexing: chunking, replication, lifecycle, multipart uploads.

Inverted indexes, sharded indexing, replication, query fan-out, ranking pipelines.

Log shipping, structured fields, aggregation, retention tiers, search-on-logs vs metrics.

Priority, idempotency, deduplication, retry policies, resource capacity allocation.

Decompose a hot counter into N shards and aggregate on read — the canonical fix for write hotspots.

Map long URLs to short ones, redirect in O(1), survive billions of clicks. The canonical first system to whiteboard — sharp scope, real scale, every primitive in play.

Generate a personalized timeline at read time, write time, or both. The pull / push / hybrid trade-off and the celebrity-fanout problem.

Video upload, encoding pipeline, CDN-backed delivery, watch-history, recommendations.

Short-form video at hyperscale. The for-you ranking pipeline is the design — recommendation, signal collection, and freshness vs personalization trade-offs.

Photo upload, feed generation, story expiration, following graph.

End-to-end-encrypted messaging at scale: presence, delivery receipts, group chats, media.

Real-time messaging with rich threads, reactions, typing indicators, and read receipts. WhatsApp's cousin with web-first reach and a different presence model.

Driver / rider matching, real-time location, surge pricing, payments, fraud detection.

Map tiles, routing on a road graph, ETA prediction, real-time traffic ingestion.

Geo-indexing, dynamic segments, search-within-radius queries at city scale.

Q&A platform: ranking answers, follow graph, notification fan-out, search.

Pluggable feed engine: ranking, freshness, deduplication, infinite scroll, caching strategy.

Real-time prefix matching: trie indexes, ranking, server-side throttling, personalization.

URL frontier, politeness, deduplication, content extraction, crawl traps, index updates.

Operational transforms vs CRDTs, presence, conflict resolution, offline edits.

Pipelines, artifact stores, environment promotion, canary / blue-green / rolling, rollback.

Idempotency, double-entry ledgers, reconciliation, gateway integration, fraud signals.

Streaming inference, KV-cache reuse, request routing, safety filters, multi-tenant GPU scheduling.

Feature stores, training data pipelines, online vs batch features, lineage, vector storage.

RAG over knowledge bases, conversation memory, escalation handoff, guardrails.

Latency-sensitive completion, repo-aware context, indexing strategies, evaluation harness.

Three-sided marketplace: restaurants, couriers, customers. Order routing, ETA, courier dispatch.

Client-side chunked replication, conflict resolution across devices, delta sync, and the LAN sync trick. Canonical 'consumer cloud storage' design.

Inventory reservation under coordinated burst load: waiting rooms, holds, atomic seat allocation, and the bot-vs-fan arms race.

Annual per-user retrospective from a year of plays. Massively-parallel batch pipeline, per-user partitioning, and the once-a-year cost shape.

A routine BGP maintenance command withdrew Facebook from the internet. Six hours of global blackout; internal tools and badge readers locked out of the buildings that hosted the fix.

Thread-limit exhaustion cascaded across services that depended on Kinesis for control-plane operations.

Why one region keeps taking down half the internet, and what 'control plane in one region' really costs.

A single WAF rule with exponential regex backtracking burned 100% CPU across every edge node simultaneously.