Complete Collection

System Design
Interview Prep

30 full system designs, a 6-phase playbook, and 157 interview Q&A pairs. Each design is a 75-minute walkthrough with architecture diagrams, deep dives, failure analysis, and the hardest questions interviewers ask.

28System Designs

6Categories

89Deep Dives

145Interview Q&As

The 6-Phase Playbook

The framework behind every design: Clarify → Estimate → High-Level → Deep Dives → Cross-Cutting → Wrap-Up. Read this first.

Common Concepts Q&A

36 foundational questions every interviewer asks: SQL vs NoSQL, caching, sharding, CAP theorem, idempotency, circuit breakers, consistency patterns, and more.

Marketplaces & Consumer Platforms

Uber Marketplace

Real-time ride-hailing: geo-spatial matching, dynamic pricing, ETA prediction, and location tracking at 500K writes/sec.

Tension: 500K location writes/sec → in-memory spatial index with matching speed vs. trip completion reliability

Facebook Social

Social graph + news feed: fan-out strategies, ranked feed generation, and the celebrity problem at 2B+ users.

Tension: 17:1 read:write ratio, celebrity fan-out → hybrid push/pull feed architecture

Amazon E-Commerce

E-commerce at scale: product catalog, cart, checkout, multi-warehouse inventory, and the browse-vs-buy consistency split.

Tension: Browsing speed vs. buying correctness → two-level inventory (soft reserve + hard commit)

Video platform: upload transcoding pipeline, adaptive bitrate streaming, CDN architecture, and recommendation engine.

Tension: Petabytes of storage + 1000:1 watch:upload ratio → cost-optimized streaming with CDN edge caching

Weather App Consumer

Weather platform: multi-source data ingestion, forecast blending & precomputation, CDN-first serving at 200K QPS, and life-safety severe weather alerts.

Tension: 100,000:1 read:write ratio with extreme cacheability → precompute + CDN-first serving + separate alert push path

Netflix Streaming

Global video streaming: Control Plane (AWS) vs. Data Plane (Open Connect CDN), per-title shot-based encoding with VMAF, ISP-embedded OCAs, and ML-driven personalization.

Tension: Browse latency (sub-200ms APIs) vs. stream quality (73 Tbps zero-rebuffer) → Control/Data plane separation with proactive CDN caching

Aggregation Platforms

Travel booking aggregator: scatter-gather search across 100+ suppliers, pricing cache, and multi-supplier booking sagas.

Tension: Search speed (<3s) vs. inventory freshness → two-phase: cached search + real-time booking

Zillow Real Estate

Real estate data platform: 110M property graph, Zestimate ML pipeline, geo-spatial search, and Premier Agent marketplace.

Tension: Every home (110M) vs. just listings (2M) → comprehensive property graph + entity resolution

Google Search Search

Web search engine: crawler, inverted index, PageRank, query serving, and ranking at billions of pages and 100K QPS.

Tension: Index freshness vs. serving speed → offline crawl/index plane + online query serving plane

Financial Systems

Robinhood Trading

Retail brokerage: order execution, real-time market data streaming, portfolio management with consumer UX + institutional correctness.

Tension: Consumer traffic patterns + institutional correctness → separated read/write paths

Bank of America Banking

Core banking: double-entry ledger, card authorization, check deposit, ACH/wire transfers, fraud detection with zero financial errors.

Tension: Real-time + batch coexistence → snapshot isolation with reconciliation as immune system

Stripe Payments

Payment processing: card network integration, double-entry ledger, PCI-scoped card vault, idempotent API design, fraud detection, and T+2 settlement pipeline.

Tension: Speed (<1s checkout) vs. correctness (every cent must balance) → sync auth + async settlement + append-only ledger

Infrastructure & Observability

Cloudflare CDN/Security

Edge network: anycast routing, CDN caching, DDoS mitigation, WAF, and Workers serverless — all with zero centralized hot path.

Tension: Zero centralized hot path → anycast + edge-local processing at 330+ PoPs

Splunk Log Search

Log analytics: schema-on-read indexing, time-bucketed inverted index, SPL query language, and 10 TB/day sustained ingestion.

Tension: Write-optimized ingestion vs. read-optimized search → time-bucketed inverted index

Datadog Monitoring

Cloud monitoring: metrics, logs, and traces ingestion, time-series storage, dashboards, and alerting across millions of hosts.

Tension: Millions of data points/sec ingestion vs. sub-second dashboard queries → TSDB architecture

Docker Containers

Container platform: Linux namespaces + cgroups for process isolation, OverlayFS for layered copy-on-write storage, content-addressable registry for image distribution.

Tension: VM-level isolation vs. container-level speed → shared kernel with defense-in-depth security layers

bit.ly URL Shortener

URL shortener: short code generation, <10ms redirect hot path, click analytics pipeline, and 100:1 read:write ratio.

Tension: Redirect speed (<10ms) vs. analytics accuracy → fast redirect path + async analytics pipeline

Developer Tools

GitHub DevTools

Code hosting: distributed Git at centralized scale, Spokes replication, pull requests, Actions CI/CD, and 200M+ repositories.

Tension: Distributed VCS → centralized hosting → Spokes replication for strong consistency

Claude Code AI Agent

Agentic coding: autonomous code generation with permission classification, tool orchestration, and the autonomy-vs-safety tension.

Tension: Autonomy vs. safety → permission classification balances speed with trust

CodeRabbit Code Review

AI code review: model cascade, ephemeral sandbox execution, precision-over-recall philosophy for PR review at scale.

Tension: Context depth vs. review latency → model cascade + ephemeral sandboxes

Model Context Protocol: JSON-RPC 2.0 wire format, host-client-server architecture, STDIO/HTTP transports, capability negotiation, and OAuth 2.1 auth.

Tension: M×N integration problem → M+N via universal protocol with host-mediated security boundary

Agent-to-Agent Protocol: Agent Cards for discovery, task-oriented lifecycle, multimodal Parts, SSE streaming, and signed-card trust model for cross-org agent collaboration.

Tension: Opaque agent interop across frameworks → task lifecycle + Agent Cards + protobuf-first multi-binding design

CrewAI Agent Framework

Multi-agent orchestration: role-based agents with backstories, Crews for autonomous collaboration, Flows for deterministic control, and the autonomy-vs-control dial.

Tension: Agent autonomy vs. production determinism → Flows (deterministic backbone) + Crews (intelligence where it matters)

LangGraph Agent Runtime

Graph-based agent runtime: Pregel/BSP execution with cyclic graphs, channel-based state with reducers, checkpointing for durability, and first-class human-in-the-loop.

Tension: Low abstraction + maximum control → Pregel super-steps give deterministic parallelism, checkpointing, and streaming for free

Identity & Security

Identity platform: SSO/SAML/OIDC federation, MFA orchestration, multi-tenant isolation, and the trust platform paradox.

Tension: Absolute security with maximum availability → SECURITY > AVAILABILITY > CONSISTENCY > LATENCY

AI Observability

Prove AI GenAI Obs

GenAI observability: trace-to-metric conversion, agentic remediation engine, self-hosted data sovereignty, and case management for nondeterministic systems.

Tension: Observing nondeterminism vs. remediating it → two-phase: telemetry foundation + agentic remediation

AI memory layer: LLM-based fact extraction, hybrid vector + graph datastore, memory consolidation (dedup/conflict resolution), and multi-tenant retrieval under 50ms.

Tension: High-quality extraction (expensive LLM calls) vs. low-latency retrieval → async write plane + sync read plane

OpenFGA / Zanzibar Authorization

Google's global authorization system: relationship-based access control (ReBAC) via tuples, graph traversal check resolution, zookie consistency protocol, Leopard indexing for nested groups, and multi-layer caching at 10M+ checks/sec.

Tension: Consistency vs. latency — zookies let apps specify freshness per-request, solving the "new enemy" problem while serving most checks from stale cache

Cursor BugBot AI Code Review

Agentic AI code reviewer at 2M+ PRs/month: V1 pipeline (8 parallel passes, randomized diff order, majority voting, validator models) → V2 agentic loop (dynamic tool use, aggressive prompting), resolution rate hill-climbing, and Autofix remediation.

Tension: Precision vs. coverage — false positives erode trust faster than false negatives, so the entire architecture optimizes for high-confidence bugs (70% resolution rate)

GenAI Architecture Reference Arch

Universal 5-layer blueprint: intake (parser flywheel), generation (composable prompts, model router), evaluation (traces, LLM-as-Judge, golden sets), 5-tier cognitive memory + knowledge graphs + GraphRAG, and adaptive orchestration.

Tension: Flexibility vs. complexity — progressive enhancement fills architecture slots incrementally; design must be complete from day one but implementation evolves week by week