System Design Study Guides

Read-heavy service that maps short slugs to long URLs at billions of lookups per day.

Hybrid fan-out — push for normal users, pull-on-read for celebrities.

Sub-50ms suggestions backed by an offline-built trie with frequency weights.

Atomic: inventory reservation, payment, order creation — with sagas for rollback.

Sell limited seats to a thundering herd without overselling or double-booking.

Bidirectional messaging with delivery + read receipts, presence, and offline push.

Async, channel-agnostic delivery with templates and rate-limiting per user.

Top-N + arbitrary-rank queries at low latency for millions of players.

Place orders against a matching engine while streaming live prices to millions of clients.

Multiple users edit the same document concurrently with conflict-free convergence.

Upload → transcode → multi-bitrate chunks → CDN-delivered HLS/DASH.

Block-level chunking, content-addressed dedup, cross-device sync.

Idempotent, auditable, double-entry ledger with strong consistency.

Crawl the web, build an inverted index, and serve ranked results in milliseconds.

Geo-shard the world; match riders to nearby drivers in milliseconds.

Spatial index over POIs answers "k nearest" queries in <100ms.

Politeness-aware distributed fetcher feeding an inverted index.

Pull-based scraping with high-cardinality TSDB and alerting pipeline.

Append-only partitioned log with at-least-once semantics and replay.

Streaming top-K over a sliding window of social signals.

Real-time inference with model versioning, A/B testing, and low-latency feature retrieval.

Empower LLMs with private data using vector embeddings and semantic search.

Turn documents (PDFs, scans, faxes, email) into trusted structured data at scale — using an LLM ensemble, deterministic auto-verification, and confidence-calibrated field-level routing to keep humans off the critical path even at ~80% single-model accuracy.

End-to-end pipeline from data ingestion to model validation and registration.

Search by image using deep feature extraction, high-dimensional indexing, and re-ranking.

Multi-modal system (text + image + video) to detect and flag policy violations in real-time.

Candidate generation (Retrieval) followed by deep ranking to optimize for click-through and watch-time.

Low-latency CTR prediction for real-time bidding and ad placement.

Real-time text completion using Sequence-to-Sequence models with ultra-low latency constraints.

Distributed serving of massive models (175B+ parameters) with low-latency streaming.

Latent Diffusion models serving thousands of concurrent generation requests.

High-throughput image processing pipeline to detect and blur faces and license plates.

Graph-based recommendation system using "friend-of-friend" signals and node embeddings.

Many-to-many language translation using Encoder-Decoder Transformers.

Massively distributed video synthesis using Spatio-Temporal Diffusion models.

Two-stage pipeline: cheap candidate generation over millions of items, then an expensive ranking model scores the shortlist for relevance.

Real-time scoring of ad candidates by predicted click probability, under a hard sub-100ms latency budget, feeding directly into an auction that decides what gets shown.

A hybrid rules-plus-model pipeline that scores every transaction in real time, balancing catching fraud against the cost of wrongly blocking legitimate customers.

The infrastructure layer underneath every other ML system in this pass — a single source of truth for features, serving both a slow offline (training) path and a fast online (serving) path, with point-in-time correctness as the hard constraint that connects them.

Token-bucket per user, enforced at the edge with Redis-backed counters.

Consistent-hash sharded in-memory store with replication and failover.

Single entry point for auth, rate-limiting, routing, and observability.

64-bit IDs: 41-bit timestamp + 10-bit machine-id + 12-bit sequence. Time-sortable.

Geo-distributed edge caches that serve static assets near the user.

Multi-layered security with OAuth/OIDC authentication, JWT authorization, network isolation, and zero-trust principles.