Discord2023
“How Discord Stores Trillions of Messages”
The same data model six years on: hot partitions, a service layer that merges identical reads, and a migration of the full history to a new database.
Live · video-processing-pipeline · stage 07 of 14
System design interview practice on real systems.
Work through a URL shortener, news feed, payment system and more. Make each design decision before you see the reasoning. Test the design against failures and changed requirements, then compare your work with the full written walkthrough. Free, no account needed.
13 systems. All of them breakable.
Each starts from requirements and the numbers. The parts the scenario gives you are already drawn; you design the rest, then it fails.
- A URL shortener like bit.ly45m
- A reliable video processing pipeline55m
- Rate limiting a public API40m
- A home timeline at 300,000 reads a second45m
- A job queue that keeps working when workers fall behind45m
- Product analytics over billions of events45m
- Notifications across email, push and in-app50m
- A payment workflow that never double-charges60m
- A look-aside cache at Facebook's scale50m
- Storing trillions of chat messages45m
- A real-time collaborative editor60m
- Live queries: screens that update in real time45m
- Sharding Postgres while it is running50m
Every system runs the same loop.
Choices are checked against a key. Written answers you mark yourself, against specific points. Nothing is a quiz for its own sake: each step changes the system you are building.
Learn
Short explanations, each with a check that answers straight away.
Decide
Choose between designs that all sound plausible.
Break it
A worker dies, a message arrives twice, a request times out.
Change it
Ten times the traffic, or a requirement from sales.
Defend it
Say what the design guarantees, and where it stops.
Built from what their engineers published.
Every system starts from a real writeup. Read theirs, then build yours and watch it break.
Stripe2020
“Stripe's payments APIs: the first ten years”
Payment APIs where a retry after a timeout must never move money twice.
Cloudflare2017
“How we built rate limiting capable of scaling to millions of domains”
Rate limits enforced in hundreds of data centres at once, where exact global counts would cost more than the attacks they stop.
Slack2017
“Scaling Slack's Job Queue”
A job queue that failed because its backlog lived in memory, and the redesign that moved the backlog to disk.
Mux2026
“What we learned from a 22-Day storage bug (and how we fixed it)”
A video API that starts playback seconds after upload by transcoding segments on demand instead of the whole file up front.
Figma2022
“Making multiplayer more reliable”
Several people editing one design at once, kept in sync by a single server per document.
Free to learn. Pay only for the AI.
Every system, concept and mock interview is free, with no account. Credits are one-time packs for AI drafts, reviews and question sets. They never expire, and a failed run gives them back.
Free
$0
forever, no account needed
Everything for learning
The whole learning product. Self-assessment is never removed or degraded.
- Every investigation, concept, company and mock interview
- Practice, progress and spaced review
- Your own projects, with the free model check
Starter
$5
one-time, in USD
15 credits
$0.33 per credit. Never expires, nothing renews.
- AI assessment of written answers, 100 a day, free
- A failed AI run returns its credits
- Unused packs refundable within 14 days
Standard
Recommended
$12
one-time, in USD
40 credits
$0.30 per credit. Never expires, nothing renews.
- AI assessment of written answers, 100 a day, free
- A failed AI run returns its credits
- Unused packs refundable within 14 days
Plus
$25
one-time, in USD
100 credits
$0.25 per credit. Never expires, nothing renews.
- AI assessment of written answers, 100 a day, free
- A failed AI run returns its credits
- Unused packs refundable within 14 days
A verified account starts with 5 welcome credits, once. Payments by Dodo Payments; prices include applicable taxes where shown.
Questions
The refunds page and the AI page have the details.
What is a credit?
A credit pays for one AI run. Each AI button shows its cost and your balance before it runs, and buying a pack shows what it pays for. New accounts get 5 welcome credits once, enough for one full deep dive: a repository draft, a review and a question set.
What happens when I run out?
AI actions stop and offer a pack instead. Everything else keeps working, and nothing is charged until you choose to buy.
What happens if the AI fails?
Its credits come back in the same response, and the page says so. That covers a provider error, a timeout, or output that fails checks after one retry. A run cut off mid-way is refunded the next time your account loads, once it is 10 minutes old.
Do credits expire?
No. Packs never expire, and nothing renews. There are no subscriptions, so there is nothing to cancel.
What is the refund policy?
An unused pack can be refunded in full within 14 days of purchase, as long as none of its credits have been used. If an AI report is wrong, the “This is wrong about my system” button refunds that run, up to 3 times in 30 days. Full terms are on the refunds page.
Is learning content free?
Yes. Investigations, concepts, companies, practice, design rounds, mock interviews, self-assessment, progress and the model check on your projects are free, with or without an account. Assessing your written answers costs no credits either, up to 20 a day (100 after any purchase).
Which AI model is used, and is my data used for training?
The models, and what is sent to them, are listed on the AI page. Providers are asked not to train on prompts and, by default, to keep no data. Repository file contents are not stored.