Amazon System Design Interview: What the Round Tests

Engineers who have passed design rounds elsewhere tend to walk into the Amazon system design interview with a method that works: requirements, estimates, boxes, a deep dive, done at minute forty-five. At Amazon that method runs into two things it was not built for. The slot is shared with behavioral questions, so there is less design time than the calendar invite suggests. And the interviewer keeps asking about things other companies leave to the end or skip: what pages you at three in the morning, what the design costs to run, what you would refuse to build.
None of that is hidden, but it is spread across Amazon's own prep pages and a lot of candidate accounts that do not always agree. This guide puts the round in order: who gets it, how the time goes, which prompts come up at which level, what the interviewer has to write afterwards, and how to prepare in three weeks.
Who gets the round, and at which level
Amazon's candidate pages for SDE II and SDE III both say, as of November 2026, to "expect at least one question on software systems design". The same sentence appears on the software development manager page, where system design also takes half of the phone screen. Below that, candidates mostly report object-oriented design instead: classes and interfaces for a parking garage or a library, not a distributed system.
The loop around the design round differs by level. Amazon describes four 55-minute interviews for SDE II and five for SDE III. Candidates describe the usual senior loop as two coding rounds, the system design round, either a low-level design round or a deep dive into a project from your resume, and the Bar Raiser. At SDE II the low-level design round and the system design round sometimes both appear and sometimes one replaces the other, depending on the team.
For Principal engineers, Amazon publishes no prep page and public accounts are thin and vary by organization. Expect design to be discussed in more than one conversation, including a system you built yourself, and expect the questions to move from "how would you build this" to "why is this the right thing to build, and what does it do to the teams around it". Treat any precise description of the Principal loop with suspicion and ask your recruiter for the round list.
How the 55 minutes go
Amazon's page for the round is brief: the interviewer "will ask questions related to your design, and you should ask questions to complete and validate your design". The rest comes from candidates, and they are consistent on the main point. The slot is 55 or 60 minutes, and a large part of it, often 20 to 30 minutes, goes to leadership principle questions before or after the design. You should plan for about half an hour of design.
That half hour tends to go like this.
Minutes 0 to 5, the prompt and your questions. The prompt is one sentence and deliberately loose. Ask who the users are, what the two or three operations that matter are, the rough scale, and what must never go wrong (lost orders, double charges, stale prices). Amazon's wording, "complete and validate", is a fair description of what the questions are for. Write the answers down where the interviewer can see them.
Minutes 5 to 15, a design that works end to end. An API or two, the main data entities, and a first diagram that handles the happy path at the stated scale. Say numbers only where they change a decision. A design that exists at minute fifteen can be improved; a perfect capacity estimate with no boxes cannot.
Minutes 15 to 28, two components in depth. Either you choose or the interviewer does. This is where data model, partition key, consistency choice, queue semantics and cache invalidation get tested. Senior candidates are expected to pick the hard part themselves and say why it is the hard part.
Minutes 28 to 35, operations and failure. What breaks first at ten times the traffic. What happens when a dependency is slow, not down. Which metrics and alarms you would ship with. How you would deploy a change to this safely. At many companies this is a closing question. At Amazon, candidates report follow-ups on failures, load and operations in nearly every account, and interviewers will pull you into them early if you do not get there.
If the behavioral questions come first and run long, you may have 25 minutes. Decide before the interview what you would cut: the estimate, never the operations.
Diagrams are drawn in an online whiteboard; Amazon's prep pages point candidates to its Bluescape tool. Try whichever tool your recruiter names before the day.
Prompts candidates report, by level
No list is official, and prompt lists on prep sites come from self-reported interviews. With that caveat, the same families keep appearing.
- SDE II. Amazon-flavoured products with a clear data model: a locker pickup system, a parking or rental system, a notification service, an order tracking page, a URL shortener, a ride-hailing or video service in outline. The bar candidates describe is working independently from a vague prompt to a sensible high-level design, with the interviewer adding load, a failure or a new feature near the end.
- SDE III. Similar surfaces with harder constraints: a ticket booking system that must not oversell, a collaborative document editor, a chat service, a rate limiter, an experimentation platform that assigns users and streams metrics. Some candidates get a "reverse" design: an existing architecture with a problem, and the job is to find where it fails. Prompts follow the team; device and satellite organizations ask about fleets of devices, firmware updates and data pipelines more than about web traffic.
- Managers. A design at the level of services and teams, such as a distributed logging system or a firmware update service, or a walkthrough of a system your own team owned and where it breaks at ten times the load.
- Principal. Often your own system, examined closely, and open questions about strategy for a whole area.
A bar raiser quoted in one prep guide draws the SDE II and SDE III line this way: the mid-level candidate assembles the core components correctly, while the senior candidate also raises security and least-privilege access, testability, how other teams would use the system without asking, and scale at the level of an organization. That matches what the questions by level guide says about the general pattern: the prompt stays the same and the expected altitude changes.
Operations and cost, the Amazon-specific probes
Amazon engineers run what they build, including the on-call rotation, and the design round reflects it. The objectives Amazon lists for the round are practicality, accuracy, efficiency, reliability, optimization and scalability, with practicality first. A candidate who draws a clean architecture and has nothing to say about running it has answered about half the question.
Have an answer ready for each of these, for whatever you design:
- Detection. The three or four metrics that tell you it is healthy, and the alarm that should wake someone.
- Blast radius. What a bad deployment or a bad input can take down, and what limits it: cells, shards, per-customer limits, a kill switch.
- Slow dependencies. Timeouts, retries with backoff, and what you serve when the dependency does not answer.
- Deployment. How a change reaches production in stages, and how it is rolled back.
- Data repair. How you would find and fix bad records after a bug, including replaying from a queue or a log.
- Cost. The component that dominates the bill, and the cheaper design you would start with if the scale were ten times smaller.
Cost deserves a sentence of its own because it is where candidates from well-funded companies get caught. "I would add a cache cluster in every region" invites the question of what that costs and whether the latency target needs it. Saying "at this scale one database with read replicas is enough, and here is the signal that tells me to move to shards" is a stronger answer than the design with every component in it.
How the principles are graded inside a design round
Each interviewer in a loop is assigned two or three leadership principles, and the design interviewer is no exception. The principles show up in two ways.
The first is explicit: behavioral questions in the same slot, often about a past technical decision. "Tell me about a design decision you regret" or "a time you simplified something" are graded like any other story, and the debrief reads them next to your other rounds.
The second is how you design. The interviewer cannot write "showed Dive Deep" without evidence, so it helps to know what the evidence looks like.
- Dive Deep. You can go two levels down on a component you named: the partition key and why, what the index looks like, what happens to in-flight messages on a consumer crash. Naming a managed service is not depth; explaining what you need from it is. The Dive Deep guide has the behavioral side.
- Think Big. You say what the design would need to become if the product succeeded, and you leave room for it without building it. At senior levels, you question whether the prompt is the right problem. See Think Big.
- Frugality. You size things to the stated scale and name the cheaper option first. See the Frugality questions.
- Customer Obsession. You begin from what the user sees and can say which failure the customer would notice and which they would not.
- Ownership. You bring up operations, security and on-call load before you are asked.
- Are Right, A Lot and Have Backbone. When the interviewer pushes on a choice, you either defend it with a reason or change it with a reason. Folding at once and refusing to move both read badly.
What the written feedback needs
After the round the interviewer writes up what happened and, by most accounts, votes before reading anyone else's notes. They need material for a few plain questions: did the candidate gather requirements, did they reach a working design, where did they go deep and was it correct, which trade-offs did they name, how did they handle pushback, and what level did it look like.
You can make that write-up easy.
- Say trade-offs as trade-offs: "I am choosing eventual consistency here because a stale count is acceptable and a failed checkout is not." A quotable sentence gets quoted.
- Label the diagram. An unlabelled box is a component the interviewer has to guess at when they write.
- When you make an assumption, say it and write it down, so a wrong assumption reads as a decision and not as a gap.
- Summarize in the last minute: what the design does, its two weakest points, and what you would do next.
- If the interviewer offers a hint, take it visibly. Prep guides built on candidate reports say interviewers also note how you respond to hints and feedback.
Differences from Google and Meta
Prep guides built on candidate reports describe Google's design round as 45 minutes of design with no behavioral questions in it; those sit in a separate Googleyness and leadership interview. The prompt is vague on purpose, interviewers differ in whether they stay high-level or go into capacity math, and the feedback goes to a hiring committee instead of a debrief of the people who met you. With no stories to tell, there is more room for depth, and some interviewers prefer that you reason from fundamentals instead of naming products. The Google software engineer guide covers that loop.
Meta, by the same kind of accounts, runs one 45-minute design round for senior engineers in one of two flavours depending on the role: system design for infrastructure roles and product architecture for product roles, the second leaning on APIs, user flows and data models. Behavioral questions again have their own round. So the practical differences at Amazon are less design time, more questions about running the system and its cost, and stories about past technical decisions graded in the same hour.
A three-week plan
Week 1, the method and the building blocks. Fix one sequence for the first fifteen minutes and practise it until it is automatic. Review the pieces you will reuse: partitioning, replication, queues and idempotency, caching, consistency choices. The system design interview guide and the worked examples for a rate limiter and a payment system cover this ground. Do three prompts alone, 30 minutes each, on a whiteboard tool.
Week 2, the Amazon layer. Redo the same three designs and add the six operational answers from the list above to each. Then write three stories about real technical decisions: one you got right against opposition, one you got wrong, one where you chose the simpler or cheaper design. Give each a number and a result.
Week 3, pressure. Do at least two full mock rounds in the real shape: 25 minutes of stories with follow-ups, then 30 minutes of design with interruptions, then a written debrief of what the interviewer could have recorded. This is the part that cannot be done alone, because the round is graded on how you respond to someone else's questions; a session with one of the coaches with Amazon on their history is the direct way to get it; ask whether they have run design rounds. Use the last days to repeat the design that went worst.
Where coaching fits
The gap between a design that is right and a design that reads well in Amazon's written feedback is hard to see from the inside. Twelve of the 40 coaches on techinterview.coach list system design as a specialty and nine have Amazon on their history, as of November 2026; the Amazon coaches page shows who covers both. A mock round with a coach from that background can be run with the behavioral questions in the same slot and written feedback in the form the debrief uses. The free practice interview has a system design round against an AI interviewer if you want a first repetition before booking anyone. It is free; sign in to start.
Related guides
- Amazon software engineer interview: the coding rounds and the rest of the SDE loop around this one.
- System design interview questions by level: how the same prompt is graded from mid-level to staff.
- Mastering system design interviews: the general method this guide assumes.
- Amazon interview process: where the design round sits and what comes after.
- Amazon phone screen interview: the stage before the loop, which for managers already includes design.
- The Amazon Bar Raiser interview: the other round where follow-ups go several levels deep.