High-Level Design

21 posts

Distributed systems, scalability, and architecture.

How Consistent Hashing Powers Distributed Systems

A deep dive into the algorithm that lets systems like Cassandra and Amazon's Dynamo route keys without reshuffling the entire keyspace when nodes come and go.

Asked at Amazon, Google, Meta +2

Jul 23, 2026·27 min read

Mediumhigh level design distributed systems networking

Load Balancing: One Front Door for a Fleet That Keeps Changing

A single server melts under Black-Friday traffic. Load balancing is how you put ten servers behind one address — and the L4-vs-L7 choice, the routing algorithm, and the health checks are where systems live or die.

Asked at Amazon, Google, Netflix +3

Aug 8, 2026·16 min read

Mediumhigh level design distributed systems caching

Caching Strategies: The Bet That Reads Repeat

A cache is a wager that the same data gets read again before it changes. Get the read/write pattern, the eviction policy, and — the hard part — invalidation right, and one box absorbs the load that would have flattened your database.

Asked at Meta, Amazon, Netflix +3

Aug 8, 2026·13 min read

Mediumhigh level design distributed systems networking

CDNs: Move the Bytes to the User, Not the User to the Bytes

A user in Sydney pays ~200ms to reach a Virginia origin before any work happens — physics you can't code around. This is how CDNs cheat the speed of light with edge PoPs, anycast routing, and HTTP caching.

Asked at Cloudflare, Akamai, Fastly +3

Aug 8, 2026·16 min read

Mediumhigh level design distributed systems microservices

The API Gateway: One Front Door for a Hundred Services

When a mobile screen fires a dozen calls to a dozen services, each doing its own auth and its own CORS, the answer isn't a smarter client — it's a single entry point that owns the cross-cutting concerns once.

Asked at Netflix, Amazon, Uber +2

Aug 8, 2026·15 min read

Mediumhigh level design distributed systems databases

SQL vs NoSQL: Model the Access Pattern, Then Pick the Store

The choice isn't a religion — it's a consequence. Design how your data is read and written first, and the datastore that fits falls out of it. Pick the store first and you'll hand-join in application code for the next three years.

Asked at Amazon, Google, Netflix +3

Aug 8, 2026·15 min read

Mediumhigh level design distributed systems databases

Database Indexing: Why B-Trees Read Fast and LSM-Trees Write Fast

A missing index turns a one-user lookup into a scan of fifty million rows. Here's how the two great index families — B+ trees and LSM-trees — buy that lookup back, and what each one charges you for it.

Asked at Amazon, Google, Meta +3

Aug 8, 2026·13 min read

Mediumhigh level design distributed systems databases

Database Replication: Copies That Disagree With Each Other

Why the same query returns different answers on two servers a second apart — and the failover, lag, and conflict traps that turn 'just add a read replica' into a 3am incident.

Asked at Amazon, Google, Meta +3

Aug 8, 2026·15 min read

Hardhigh level design distributed systems databases

Database Sharding: When One Machine Stops Being Enough

Your 4 TB users table no longer fits — or no longer keeps up with writes — on a single node. Sharding is the answer, and the shard key you pick quietly decides everything that follows.

Asked at Amazon, Google, Meta +3

Aug 8, 2026·16 min read

Hardhigh level design distributed systems databases storage

Storage Engines: LSM-Trees vs B-Trees

Every database makes one foundational choice — how to lay bytes on disk. B-trees update in place; LSM-trees never overwrite, they append and compact. That single decision sets your write throughput, read latency, and space usage. Drive an LSM write path — memtable, flush, compaction — and see the tradeoff for yourself.

Asked at Google, Meta, Amazon +3

Aug 9, 2026·14 min read

Mediumhigh level design distributed systems messaging

Message Queues & Event Streaming: Stop Making Users Wait for Work They Don't Care About

A checkout that emails, bills, and updates inventory before it answers is one flaky partner away from a timeout. How queues and logs let producers and consumers stop holding each other hostage.

Asked at Amazon, LinkedIn, Netflix +3

Aug 8, 2026·17 min read

Mediumhigh level design distributed systems networking

Rate Limiting: How a Bucket of Tokens Keeps Your API Alive

One misbehaving client with a retry loop can take down a service faster than any DDoS. Rate limiting is the cheap, boring control that decides who gets served — here's how the five classic algorithms actually work, and how to make one correct across a fleet.

Asked at Stripe, Cloudflare, GitHub +2

Aug 8, 2026·16 min read

Mediumhigh level design distributed systems networking

Real-Time Updates: Polling, Long Polling, SSE, and WebSockets

HTTP can't push — the server only answers when asked. Here's the ladder of techniques that fake or fix that, each one solving the last one's flaw, and the fan-out problem that breaks all of them at scale.

Asked at Slack, Discord, WhatsApp +3

Aug 8, 2026·16 min read

Mediumhigh level design distributed systems consistency

CAP Theorem: The Choice You Only Make When the Network Breaks

Everyone quotes 'pick 2 of 3' and everyone is wrong. CAP is a decision forced only during a partition — and PACELC is the sharper lens for the 99.9% of the time when the network is fine.

Asked at Amazon, Google, Microsoft +3

Aug 8, 2026·12 min read

Hardhigh level design distributed systems consensus

Consensus & Leader Election: How a Cluster Agrees on One Truth

Split-brain corrupts data quietly. This is how Raft gets N crash-prone nodes to agree on a leader, a log, and a lock — and why a majority is the only quorum that can't lie to itself.

Asked at Google, Amazon, Microsoft +3

Aug 8, 2026·17 min read

Hardhigh level design distributed systems consensus

Raft, Driven: Watch a Cluster Elect, Replicate, and Heal

Everyone can recite 'Raft elects a leader by majority.' The hard part is the log — how a follower that missed writes during a partition gets repaired, and the one subtle rule that stops a committed entry from ever being lost. Drive a live 5-node cluster, then take it apart.

Asked at Google, Amazon, Microsoft +4

Aug 9, 2026·20 min read

Hardhigh level design distributed systems transactions

Distributed Transactions: When One Business Action Spans Many Services

The payment succeeded but the inventory reservation failed — and there's no ROLLBACK that spans both. How 2PC, Sagas, and the Outbox pattern give you atomicity when a single ACID transaction is no longer on the table.

Asked at Amazon, Uber, DoorDash +3

Aug 8, 2026·17 min read

Mediumhigh level design distributed systems reliability

Idempotency: How to Retry a Payment Without Charging Twice

The network fails after your server does the work but before the answer gets home, so the client retries — and a naive charge endpoint bills the card twice. Idempotency is the discipline that makes every retry safe.

Asked at Stripe, Amazon, PayPal +3

Aug 8, 2026·15 min read

Mediumhigh level design distributed systems data structures

Bloom Filters: The Yes/No That Saves a Disk Seek

A tiny bit-array that answers 'definitely not here' or 'maybe here' in constant time — how Cassandra, HBase, and Chrome skip millions of expensive lookups for keys that were never there.

Asked at Google, Amazon, Meta +3

Aug 8, 2026·12 min read

System Design Case Studies2-part series

Hardsystem design interview distributed systems message queues realtime

Design a Chat System

WhatsApp-scale messaging, built end to end: persistent WebSocket gateways, a user→server registry, store-and-forward for offline devices, the one-tick / two-tick / blue-tick delivery state machine, message ordering, and group fan-out. Drive the delivery protocol yourself, then take it apart.

Asked at Meta, WhatsApp, Slack +3

Aug 9, 2026·14 min read

Hardsystem design interview distributed systems caching scalability

Design a News Feed

The Twitter/Instagram timeline, worked end to end. The whole system pivots on one decision — do you build a follower's feed when someone posts (push), or when they open the app (pull)? Toggle between them, watch the write and read cost flip, and see why one celebrity breaks the naive answer.

Asked at Meta, Twitter, Instagram +3

Aug 9, 2026·13 min read