Amazon system design interview

Round 5 of 7 · 57 questions reported

See the whole loop
Length
55 min
Format
Design discussion
They grade
Practicality, reliability and scalability, after clarifying questions
Heads-up
Applied scientist and MLE loops get ML system design instead.

Questions Amazon asks in this round 57

  • Transformer and attention fundamentalsML designExplain how a transformer works from the inside: how self-attention is computed, why the scores are scaled, how encoder and decoder masking differ, and what multiple heads buy…
    Last asked Sep 2026
  • ML fundamentals deep dive for ML rolesML designAfter a short walk through your resume, expect fast questions on machine learning basics, usually centred on your own area (language models, vision, recommendations). For each…
    Last asked Sep 2026
  • Design a distributed job schedulerSystem designDesign a service that runs millions of scheduled and on-demand jobs (cron-like and one-off) reliably across a fleet of workers.
    Last asked Sep 2026
  • Design a distribution-center inventory update serviceSystem designDistribution centers send stock updates. Design the service that keeps the latest quantity, availability and last-update time for every location and item pair and answers lookups…
    Last asked Sep 2026
  • Decoder-only LLM architecture and how vLLM serves itML designExplain how a decoder-only language model is built, from tokens through the stack of attention and feed-forward layers to predicting the next token. Then explain how an inference…
    Last asked Sep 2026
  • Design a product image upload pipeline for sellersSystem designDesign the pipeline sellers use to upload product images: accepting the upload, processing the images, storing them, and serving them on product pages.
    Last asked Aug 2026
  • Design faceted product searchSystem designDesign product search for a large catalog where shoppers narrow results with facets (filters on product attributes) and see how many results each filter value has. Cover indexing…
    Last asked Aug 2026
  • Design a platform to train and serve ML modelsML designDesign a platform that lets teams train machine learning models and serve them in production. Cover data access, training jobs, storing and versioning models, deployment, and…
    Last asked Aug 2026
  • Design a metrics monitoring systemSystem designCollect metrics from every host and service, store them as time series, power dashboards over months of data, and alert when thresholds are crossed.
    Last asked Jul 2026
  • Tokenizer design and KL-regularized fine-tuningML designDiscuss how you would design a tokenizer for a language model and the trade-offs involved. Then explain supervised fine-tuning with a KL penalty that keeps the model close to its…
    Last asked Jun 2026
  • Design a distributed web crawlerSystem designCrawl billions of web pages politely and repeatedly, deduplicating URLs and content, to feed a search index or a training-data pipeline.
    Last asked May 2026
  • Design SpotifySystem designStream music to hundreds of millions of listeners: catalog, search, playlists, smooth playback over bad networks, and offline downloads.
    Last asked May 2026
  • Design UberSystem designRiders request trips, nearby drivers get matched, and both see live location until drop-off, with pricing and payment, for millions of trips a day.
    Last asked May 2026
  • Design a highly available load balancerSystem designDesign a load balancer that stays up when its own machines fail: how traffic is spread across backends, how unhealthy backends are detected, and how the load balancer itself…
    Last asked May 2026
  • Web-scale crawl cleaning for pretraining dataSystem designDesign a staged batch pipeline that turns raw web crawl into a deduplicated, quality-filtered, decontaminated pretraining corpus on a fixed cluster and deadline. Reported on…
    Last asked Mar 2026
  • Design a multi-provider financial dashboardSystem designDesign a dashboard that shows each user financial metrics such as account balances and stock holdings. The data comes from several external providers, such as stock data feeds…
    Last asked Feb 2026
  • Design a computer-use agentAI systemDesign an agent that operates a computer the way a person does: it looks at the screen, decides the next action and clicks or types until a task is done. Cover how the model…
    Last asked Jan 2026
  • Critique a design doc from the team's backlogSystem designThe interviewer hands over a real design doc and asks for a critique: surface risks, propose alternatives, and say what is missing before this could ship.
    Last asked Jan 2026
  • Debug a model that does worse online than offlineML designA model scored well in offline evaluation, but after launch its online metrics are worse. How do you find out why? Walk through the likely causes (training-serving skew in…
    Last asked Jan 2026
  • Design a globally consistent metadata serviceSystem designDesign a metadata service that stays strongly consistent across regions: every read reflects the latest committed write wherever it is served. Cover the data model, how writes are…
    Last asked Jan 2026
  • Design a personalized search ranking systemML designRank search results for each user's query and context so they find and book (or order) what they want, learning from clicks and conversions.
    Last asked Jan 2026
  • Design an LLM output quality validation systemAI systemDesign a system that checks the quality of a language model's answers before and after they reach users. Cover what "quality" means for the product, automatic checks and…
    Last asked Dec 2025
  • Design an S3-like object storage systemSystem designStore trillions of objects of any size with eleven nines of durability, and serve reads and writes over an HTTP API at high throughput.
    Last asked Dec 2025
  • Floating-point formats for training and an ablation studyML designCompare the floating-point formats used to train and run models (FP32, FP16, BF16, FP8 and others): range, precision, speed and when each is a good choice. How would you detect…
    Last asked Dec 2025
  • Tensor parallelism versus pipeline parallelismML designA transformer is too big for one accelerator. Explain how tensor parallelism and pipeline parallelism each split it: what is cut, what crosses the wire and when, where the idle…
    Last asked Dec 2025
  • Design a logo infringement detection systemML designDesign a machine-learning system that finds listings or images that misuse a brand's logo. Cover how brands register their logos, how you detect and match logos in new images at…
    Last asked Nov 2025
  • Design an online Minesweeper platformSystem designDesign a platform where people play Minesweeper online. Cover generating boards, the API for revealing and flagging cells, where game state lives and who is trusted to know mine…
    Last asked Nov 2025
  • Design InstagramSystem designUsers post photos and videos, follow each other and scroll a feed; design upload, storage, feed generation and serving media worldwide.
    Last asked Oct 2025
  • Design a multimodal neural networkML designDesign a neural network that takes several kinds of input at once, such as text and images. Explain how each input is encoded, where and how the modalities are fused (early, late…
    Possibly retired
  • Design a risk management ticketing systemSystem designDesign a ticketing system that a security and risk team uses to track risks: open a ticket, assign an owner, set severity and deadlines, escalate when it is overdue, and close it…
    Possibly retired
  • Design TinyURLSystem designTurn long URLs into short links and redirect visitors fast, at billions of redirects a day, with optional custom aliases and expiry.
    Possibly retired
  • Price rides with fairness and poolingSystem designWrite and design the pricing logic of a ride-hailing service: compute a fare from distance, time and demand, keep prices fair between riders, and price pooled rides where several…
    Possibly retired
  • Design a GitHub-like developer communitySystem designDesign the user-facing community side of a GitHub-like site: profiles, following people and repositories, stars, activity feeds and notifications. The interviewer stressed very…
    Possibly retired
  • Design a feature flag serviceSystem designLet teams turn features on for a percentage of users or specific accounts, with changes reaching every service within seconds and no per-request network call.
    Possibly retired
  • Design DropboxSystem designSync files across a user's devices and shared folders: upload large files reliably, transfer only what changed, and handle conflicting edits.
    Possibly retired
  • Design an ad click aggregation systemSystem designCount ad clicks per ad and per time window from billions of events a day, accurately enough to bill advertisers, with results available within minutes.
    Possibly retired
  • Design an online bookstoreSystem designBrowse and search a catalog, keep a cart, check out and track orders, without overselling limited stock.
    Possibly retired
  • End-to-end ML coding: from raw data to predictionsML designA hiring-manager problem-solving round: given a simplified dataset for a business problem, write the whole pipeline in code. Clean and process the data, engineer features…
    Possibly retired
  • Design a bot detection systemML designSeparate automated traffic from real users on a large website, in real time, while bots adapt to whatever you deploy.
  • Design a distributed key-value storeSystem designStore and serve billions of keys with low latency, partitioned across machines and replicated to survive failures, with tunable consistency.
  • Design a driver location heatmapSystem designShow near-real-time density of available drivers and rider demand on a city map, updated every few seconds from millions of location pings.
  • Design a hotel booking systemSystem designSearch hotels by place and dates, show availability and prices, and book rooms without double-booking, including holds and cancellations.
  • Design a quota serviceSystem designTrack and enforce how much of each resource (API calls, storage, compute) every tenant may use, with fast checks and limits that stay accurate across regions.
  • Design a rate limiterSystem designDesign a service that limits how many requests each API client can make. It sits in front of every API endpoint at high traffic.
  • Design a real-time air traffic control systemSystem designTrack every aircraft in a region from radar and transponder feeds, predict conflicts, and alert controllers within a second, with no single point of failure.
  • Design a VM bandwidth rate limiterSystem designCap the network bandwidth each virtual machine in a cloud region may use, enforcing per-tenant limits across many hosts without adding latency to packets.
  • Design an address autofill systemML designAs a user types an address, suggest the full, valid address they mean, ranked by likelihood and personalized by location and history.
  • Design ChatGPTAI systemDesign a consumer chat product backed by a large language model: conversations, streamed replies, history across devices, at 100M weekly users.
  • Design typeahead suggestionsSystem designSuggest completions as the user types into a search box, within tens of milliseconds, ranked by popularity and fresh with trending queries.
  • Design YouTubeSystem designUpload videos, transcode them into many formats, and stream them worldwide with adaptive bitrate, along with views, likes and comments.