Post
System Design Practice Problems, From Easy to Hard
41 system design practice problems ordered from easy to hard, from a URL shortener to a distributed database, each with what it teaches and a link to practise it with graded feedback.
By Codemia • November 12, 2026
5 min read
Most lists of system design questions are a wall of 100 titles in no particular order. That is not much help when you are deciding what to practise next. This list has 41 problems from the Codemia system design catalogue, ordered from easy to hard, with one line on what each problem actually teaches.
Every problem links to its practice page, where you can write your own design, draw it on a whiteboard and get it graded. Problems marked (free) can be practised without a paid plan.
Pick a problem, set a 45 minute timer, and write the design before you read any solution. Requirements first, then rough numbers, then the high level design, then one or two deep dives. Only compare against the editorial once you have committed to your own answer.
How the difficulty levels work
The difficulty is about how many hard decisions a problem forces, not how famous the company is.
- Easy problems have one central idea (generate a short key, store a tree, count fast). Good for learning the shape of an answer.
- Medium problems add scale or contention: fan-out, rate limits, many writers on one record. Most mid-level interview questions sit here.
- Hard problems combine several of those ideas, and the interviewer will push on two or three deep dives. Typical of senior loops.
- Advanced problems ask you to design the infrastructure that the other answers take for granted: the database, the DNS, the scheduler. Useful for staff-level and infrastructure roles.
If you are new to system design, start at the top and do two or three problems per level before moving on. If you have an interview soon, skip to the level of your target role and work through it in order.
Easy: one core idea each
- URL shortener (TinyURL) (free). Key generation, a read-heavy store, redirects and caching the hot links.
- Pastebin. Blob storage next to a metadata database, expiry, and why the text does not belong in the row.
- Parking lot (free). Modelling entities and state, and handling two cars claiming the same spot.
- Nested comments. Storing a tree in a database: adjacency lists, materialised paths, and paging through replies.
- Video view counter. Counting at high write rates: batching, sharded counters, and how exact the number needs to be.
- Online presence indicator. Heartbeats, timeouts, and fanning a status change out to the people who care.
- Conference room booking. Preventing double bookings with constraints or locks, and time ranges as data.
- Craigslist. A classic listings site: CRUD at scale, search by category and city, image uploads.
Medium: scale and contention
- Twitter (free). Timelines with fan-out on write versus fan-out on read, and the celebrity problem.
- Instagram. Media upload pipeline, CDN delivery, and a feed built from people you follow.
- Typeahead suggestion. Prefix tries, precomputed top results per prefix, and latency budgets per keystroke.
- API rate limiter. Token bucket and sliding window, and keeping counts consistent across many servers.
- Distributed unique ID generator. Snowflake-style IDs, clock skew, and sortable versus random keys.
- Web cache. Eviction policies, invalidation, and what happens to the database when the cache goes cold.
- Task scheduler. Delayed jobs, at-least-once execution, and leases so two workers never run the same task.
- Webhook notification service. Retries with backoff, idempotency keys, and isolating one slow customer from the rest.
- Facebook Messenger. Persistent connections, message ordering, delivery receipts and offline delivery.
- YouTube or Netflix. Transcoding pipelines, adaptive bitrate streaming, and CDNs doing most of the work.
- Digital wallet. Ledgers, double-entry balances, and transfers that never lose or create money.
- Flash sale. Thousands of buyers for a hundred items: queues, inventory reservation and overselling.
- Top-K request analysis. Heavy hitters in a stream: count-min sketch, windows, and exact versus approximate answers.
- Live comments. Pushing comments to viewers in real time with WebSockets or server-sent events.
- Distributed lock. Leases, fencing tokens, and why a lock in Redis is weaker than it looks.
Hard: senior loop favourites
- Ticketmaster (free). Seat holds that expire, a virtual waiting room, and contention on the most popular events.
- Dropbox (free). Chunked uploads, deduplication, sync across devices, and resolving conflicting edits.
- TikTok (free). A recommendation-driven feed, video storage and prefetching the next clip.
- ChatGPT (free). Streaming tokens, GPU capacity as the bottleneck, and conversation storage.
- Facebook News Feed. Ranking, a hybrid fan-out model, and keeping the feed fresh without rebuilding it per request.
- Google Docs. Real-time collaborative editing with operational transforms or CRDTs, and cursor presence.
- Uber backend. Geospatial indexing, matching riders to drivers, and location updates every few seconds.
- Key-value store. Partitioning, replication, quorums and what a node does when it rejoins.
- Web crawler. URL frontier, politeness per host, deduplication and recrawl scheduling.
- Ad click aggregation. Stream processing with exactly-once counting, late events, and reconciliation with batch.
- Metrics monitoring and alerting. Time-series storage, downsampling, and alert rules that do not page on noise.
- Distributed message queue. Partitioned logs, consumer groups, retention and replay, much like Kafka.
- LLM inference service. Batching requests on GPUs, KV cache memory, and routing across model replicas.
Advanced: infrastructure and staff-level problems
- Distributed OLTP database. Distributed transactions, consensus per shard, and serialisable isolation across regions.
- Domain Name System. Hierarchical resolution, caching with TTLs, and anycast for a service everyone depends on.
- Container orchestration. A control plane in the style of Kubernetes: desired state, scheduling and reconciliation loops.
- GPU cluster scheduler for ML training. Gang scheduling, fairness between teams, and preemption of long training jobs.
- Big data processing pipeline. Batch and stream layers, partitioned storage, and backfills that do not take the system down.
A four week plan using this list
The full 4-week plan explains each week in detail. In short:
- Week 1: the eight easy problems. Focus on the order of your answer: requirements, estimates, API, data model, high level design.
- Week 2: eight of the medium problems, picked across different areas (one feed, one queue, one counter, one payment).
- Week 3: the rest of the medium problems, plus two hard ones. Start practising deep dives out loud.
- Week 4: hard problems only, under a strict timer, ideally with a mock interview at least twice.
Redo any problem where the grading said you missed a key point, a few days later, without looking at your first attempt.
Where to go next
- The full catalogue has 120+ system design problems, plus object-oriented design, ML system design and agentic AI design problems: codemia.io/system-design.
- New to system design? Start with System Design Explained: a beginner's guide.
- Want to rehearse out loud? Compare AI, peer and paid mock interviews.
- Not sure which platform to practise on? See the best system design interview practice platforms in 2026.
- For a worked example of a full answer, read the metrics monitoring and alerting walkthrough.