139. Design Similar Listings for a Vacation Rental Platform
Pick 12 homes a guest could book instead, from listing embeddings learned on browsing sessions, filtered by dates and party size, then ranked.
217 worked designs. 204 are answered at junior, senior and staff level.
Company tags are community-reported. Counts on cards show how many people reported that design.
Pick 12 homes a guest could book instead, from listing embeddings learned on browsing sessions, filtered by dates and party size, then ranked.
Suggest people a member knows out of a billion: bounded friends of friends, affiliations and contacts, two ranking heads, GNN embeddings, privacy and abuse.
Order each home feed so the time is worth it: candidate sources, a multi-task ranker and value model, integrity re-ranking and feedback loops.
Rank upcoming events when every event is new and expires: geo and time candidates, live features, calibrated ranking, a two-sided market.
Find the right videos for a typed query among billions: BM25 and a dual encoder over text, speech and frames, LambdaMART on debiased clicks, human raters.
Selfies in, professional headshots out: per-order LoRA on SDXL, an identity-encoder preview, face-match ranking, GPU pools, consent and deletion.
Images at 2K–4K without paying for every pixel: latent diffusion, an SR cascade with noise augmentation, tiled decoding, distilled drafts, cost per megapixel: one-hour boards for junior, senior and staff, with the theory.
Short clips from a prompt: latent video diffusion, recaptioned data, spacetime attention, cascades, multi-GPU jobs, previews, fair queues, provenance.
Blur every face and licence plate in billions of street panoramas: tiles for tiny faces, recall-first detection, a batch pipeline that fails closed.
The URL shortener with a payload: the same id generation and cache-first read path, but the value is kilobytes of text, so it moves out of the database and into object storage.
A cloud-native relational database like Aurora: the log is the database, six copies in three AZs with 4/6 quorums, shared-storage replicas, Global Database.
Every distributed-systems fundamental in one place, each with a figure, applied to a key-value store.
Run strangers' code on shared hosts in Firecracker microVMs, scale each function from zero in seconds, load 10 GB images lazily and bill by the millisecond.
Push fan-out to queues, functions, webhooks, inboxes and phones: stored in three AZs before the ack, retries per protocol, DLQs, filters, FIFO order.
A consistent key-value store like etcd: Raft across three AZs, MVCC revisions in bbolt, watches, leases and compare-and-swap, run as a fleet behind Kubernetes.
Virtual machines launched in seconds onto shared servers: zonal control planes, bin packing, Nitro offload cards, a mapped virtual network and Spot.
SQL tables split into Paxos groups across regions, ordered by TrueTime and commit wait, and read without locks at any replica.
An LSM-tree key-value engine inside its caller: a WAL and a memtable, levelled compaction, filters and caches, then one engine per node under Raft.
Shards of fixed capacity, partition keys hashed to ranges, three copies in three AZs, live resharding, polling and push readers with KCL leases.
Layer-7 proxies and layer-4 flow hashing behind DNS, health checks and draining, a control plane outside the request path, cells and zonal shift.
DNS steering to 750 POPs, regional caches and Origin Shield with request collapsing, purges in seconds, edge functions, DDoS at the edge.
Authoritative DNS at cloud scale: anycast stripes, shuffle-sharded name servers, routing policies, health-checked failover, a control plane that can fail.
A sparse, sorted, versioned map cut into tablets: stateless tablet servers over S3, a replicated log, a lock service and a master.
Isolated tenant networks on one shared fabric: an overlay keyed by VPC id, rules on every host's Nitro card, a mapping service per AZ, gateways as fleets.
A state-machine orchestrator: steps that happen once, waits of up to a year, retries with jitter, .sync jobs and task tokens, Distributed Map over 100 M items: one-hour boards for junior, senior and staff, with the theory.
Delivering LLM tokens from model servers to millions of apps: SSE through CloudFront and an ALB, resumable streams and a stop that frees the GPU. One-hour boards for junior, senior and staff, with the theory behind them.
Llama 3.1 70B with prefill and decode on separate GPUs and the KV cache handed over NVLink or EFA, two pools sized from the token mix.
A 3 B LLM on phones and laptops: bandwidth math, 4- and 2-bit weights, LoRA per feature, NPU runtimes, cohort rollouts, a private server.
Which model answers each request: rules, a score per tier, thresholds fitted on judged labels, cascades where a check exists, and the loop that keeps quality measured: one-hour boards for junior, senior and staff, with the theory.
Text in, unit vectors out for every product at 100,000 requests a second on AWS GPUs, batched by tokens and never mixing model versions.