System Design (HLD)

A repeatable 7-step framework so you never freeze: scope → estimate → API → data → high-level → deep-dive → trade-offs. Plus the back-of-envelope numbers worth memorizing.

Where to cache, what to cache, how to invalidate, and the failure modes (stampede, penetration, avalanche) interviewers love to probe.

Vertical vs horizontal scaling, why stateless services are the unlock, load-balancer types and algorithms, and how to handle sessions when any server can take any request.

What CAP actually forces you to choose during a partition, the spectrum of consistency models, and when strong consistency is worth its latency cost.

SQL vs NoSQL with a real reason, leader-follower replication, the four sharding strategies, and the hotspot/rebalancing traps that consistent hashing solves.

Why you push slow work off the request path, queue vs pub/sub, the three delivery guarantees, and why idempotency is the price of reliability.

The five algorithms (fixed window, sliding log, sliding counter, token bucket, leaky bucket), how to make them work across many servers with Redis, and the trade-offs.

A full HLD walkthrough using StockStump: design a system where asset prices move with live events and thousands of users trade in-play — covering fan-out, the concurrency invariant, and latency-arbitrage fairness.

A full HLD walkthrough using LandAI: pull data from third-party sources legally and reliably, wrap every record in provenance, and never present a guess as a fact.

The messaging system design: persistent connections, the online/offline delivery split, message ordering, receipts, and group fan-out.

Video streaming at planet scale: the transcode pipeline, why the CDN is 95% of the design, adaptive bitrate, and the metadata/play split.

The feed problem: fan-out on write vs read, the celebrity exception, photo storage done right, and why the home feed is precomputed.

Netflix's problem plus user-generated chaos: the upload pipeline, view counting at scale, long-tail caching, and comments under fire.

The geospatial one: indexing moving drivers, the matching flow, trip state machines, and why this design is about location, not content.

The canonical fan-out interview: hybrid timelines under a power law, retweets, trending topics as a streaming problem, and search over everything ever said.

The sync problem: chunking and content hashes, delta sync, conflict resolution, sharing permissions, and why metadata and blocks live apart.

Deliver events to merchants' URLs reliably: at-least-once delivery, per-merchant ordering, an append-only event store, retries with a dead-letter queue, Kafka for async fan-out, per-merchant rate limiting, signing, and the observability that proves it works.

The same app at six orders of magnitude — what breaks at each stage, what you add, and what each addition costs. The capstone mental model for all of HLD.