Start with a template. Work through each step. Ask Coach when you need a second opinion.
Company tags are community-reported. Counts on cards show how many people reported that design.
An S3-inspired object store: bytes before metadata, a partitioned namespace with cache coherence, then repair, heat, lifecycle and safe change. Three interview boards with explicit assumptions and AWS source boundaries.
Build a queue around visibility leases and explicit acknowledgement, then add replicated partitions, FIFO groups, tenant fairness and bounded redrive: three self-contained interview boards.
Radius, k-nearest and map-box search over places that rarely move and drivers that ping every few seconds: R-trees, geohash, H3 and S2 cells, Redis GEO and OpenSearch. One-hour boards for junior, senior and staff, with the theory.
Unique visitors, top pages, seen URLs and p99 latency over 10 B events a day with HyperLogLog, Count-Min, Bloom filters and t-digests.
Nearest-neighbour search over a billion embeddings: HNSW, IVF-PQ and disk graphs, filters, segments, sharding, and hybrid search with re-ranking.
The database under a metrics platform: a WAL and head block, compressed chunks, a label index, rollups, compaction and blocks in S3.
Every committed change in Aurora and DynamoDB, read from the log and delivered in order per row to search, caches, Redshift and an S3 lake.
Deny requests by caller address, range or URL at a global edge, with lists from governments, threat feeds and detection reaching 300 proxies in seconds.
Nights as counted inventory, holds that never double-book, search that knows availability, and payments that reconcile.
Dish reviews only from customers who ordered them, votes counted from a change stream, reviews ranked by the Wilson bound, payouts made exactly once.
Ten charities, three days, $100 M: charge through a third-party processor exactly once, queue the broadcast spikes, keep live totals on sharded counters. One-hour boards for junior, senior and staff, with the theory behind them.
Buy the cheapest offer under a maximum price from 1,000 sellers: cached offers, coalesced quotes, a hold-order-capture saga, open requests.
Turn 10 pushes a second into job graphs on ephemeral runners, with fair queues per pool, leases fenced by attempt and live logs.
Real-time rated chess with server-validated moves, lag-compensated clocks, pairing by rating and a million watchers on one game.
Rank 200 million players in real time with sharded Redis sorted sets, count every game exactly once, and close a season fairly.
Count 4 million moving drivers per map cell from a million pings a second, and serve the density to a million viewers as cached map tiles.
Design a team chat app like Slack: channels and DMs in real time, threads, unread counts and search. One-hour boards for junior, senior and staff, with the theory behind them.
Find restaurants that deliver to you, check out without double charges, time couriers to the kitchen and track them live.
Ten terabytes of logs an hour from host agents through Kafka into tiered OpenSearch and S3, with regex search, live tail and exceptions grouped into issues.
Record every money movement as balanced, append-only entries, exactly once, with holds, payouts and reconciliation.
One set of tags across Jira issues, Confluence pages and Bitbucket pull requests: batch renders, tag pages that never leak, suggestions, popular tags.
Billions of files, hundreds of petabytes, one strongly consistent tree: a namespace partitioned by directory, chunk servers, replication, repair, erasure coding. One-hour boards for junior, senior and staff, with the theory behind them.
Signed HTTPS callbacks for a billion events a day, at least once, retried for three days, with no endpoint able to slow another.
Thousands of GPUs shared by many teams: gang admission, topology-aware placement, quotas with lending, checkpointed preemption and failure recovery.
A 500 GB model on 1,000 GPU servers in minutes: chunks and hashes, a topology-aware swarm, signed manifests, bandwidth budgets and waves.
Define a feature once, train on point-in-time correct history and serve 200 fresh values in under 10 ms.
Train 1 B to 400 B models on 8 to 10,240 H100s: DDP, ZeRO and FSDP, tensor and pipeline parallel over NVLink and EFA, checkpoints, hot spares, MFU and goodput.
A chat assistant on models we train and serve: next-token framing, three training stages, resumable SSE streams, prefix caching, token quotas, safety, evaluation.
From registry to traffic for hundreds of models: gates, shadow and canary with automatic rollback, packed CPU and GPU pools, LLM serving, cells and audit.
Open-weight LLMs from 8B to 405B served to many products over a streaming OpenAI-style API on AWS GPUs.