The Prompt Bank — a system design interview rehearsal surface
Twenty system design interview prompts stated the way an interviewer opens, each with clarifying questions, the numbers to ask for, a strong-answer outline, and a failure sweep — hidden until you’ve attempted the answer. Plus the 45-minute mock protocol and a self-grading rubric.
Data infrastructure
The storage, streaming, and query systems a data-platform loop leans on. Object stores, logs, exactly-once counting, and the machinery that moves rows around a cluster.
- Design a transaction log that turns a plain bucket of files in object storage into a table people can safely read and write at the same time.
- Design a query engine that runs thousands of customers’ SQL on one shared fleet of machines, fairly, without one tenant starving the rest.
- Design a stream processor that counts events in real time and never double-counts, even when a machine crashes and restarts mid-flight.
- Design a message queue that lets services hand work to each other, never loses a message, and survives a consumer that can’t keep up.
- Design a pipeline that ingests billions of measurements a day and serves fast dashboards and alerts over them.
- Design a scheduler that runs pipelines of dependent jobs so no crash loses a run and no overlap runs the same job twice.
- Design an object storage service that makes losing a file a once-in-ten-million-years event, even when a whole data centre burns down.
- Design a distributed cache that stays fast when one celebrity key suddenly pulls three hundred thousand requests a second.
- Design a system that answers “what are the top-k right now” over a firehose of events — trending posts, heavy hitters, a live leaderboard.
Classic product systems
The canon of product-side design questions. Feeds, chat, rate limits, keys, and location — the ones almost every generalist loop pulls from.
- Design a news feed: when someone opens the app, show a fresh, ranked list of posts from everyone they follow.
- Design a chat system: one-to-one and group messages, delivered in order, that reach people whether they’re online or not.
- Design a rate limiter that caps each client to N requests per window, fairly, across a fleet of servers.
- Design a key-value store that keeps the promise saved data survives a crash and stays available during a network partition.
- Design a service that hands out unique IDs at scale, roughly time-ordered, with no central coordinator on the hot path.
- Design a proximity service: given my location, quickly return the nearby things — restaurants, drivers, friends.
AI infrastructure
The serving, retrieval, and integration systems behind ML and agent products — the ones an AI-platform loop probes once the model itself is a given.
- Design a platform that serves a large language model to many users at once, cheaply, without one long request starving the rest.
- Design a retrieval-augmented generation system: answer questions grounded in a private corpus, with fresh, relevant context.
- Design a feature store: compute features once and serve the exact same values to model training and to low-latency online inference.
- Design a platform that lets an LLM agent call external tools safely — pick a tool, run it, feed the result back — at scale.
- Design a webhook delivery system: reliably send event notifications to thousands of customer endpoints, retrying failures without duplicating.
Machine learning systems
The applied-ML design canon — retrieval, ranking, moderation, and recommendation — plus the method that unifies them. The rounds where the graded skill is turning a fuzzy product goal into a crisp ML objective.
- Design visual search: a user uploads a photo of a product and gets back the most visually similar items in the catalog, in milliseconds.
- Design a click-through-rate ranking system: order candidate ads by how likely each is to be clicked — and make the probabilities honest, not just the order.
- Design a harmful-content moderation system: flag policy-violating posts at scale, auto-block the clear cases, and route the uncertain ones to a bounded human-review queue.
- Design a video recommendation system: pick the next handful of videos a user is most likely to enjoy from a catalog of millions, in under a hundred milliseconds.
- You’re handed “design an ML system for X” with forty-five minutes — show the method, not just an architecture.
Generative AI systems
Serving and grounding the generative stack — LLM assistants, retrieval-grounded answers, diffusion image services, and inline latency-critical suggestions. Where the model is a given and the system around it is the design.
- Design an assistant chatbot: serve a large language model to millions of users in multi-turn conversations, cheaply, with tokens streaming back fast.
- Design a document-grounded assistant that answers only from a private corpus, cites its sources, and refuses when it doesn’t actually know.
- Design a text-to-image service: turn a prompt into an image at scale on GPUs, with a predictable cost per image and a queue that never melts.
- Design Gmail-style smart compose: finish the user’s sentence inline as they type, fast enough that the suggestion beats their next keystroke.