← All posts
general

Google System Design Interview: Format and How to Prepare

Engineers who have passed design rounds elsewhere often fail the Google system design interview with an answer they were proud of. The usual reason is that they drew the architecture they would build at work: a managed queue here, a managed database there, a load balancer in front. At Google the interviewer is reported to ask what is inside the box. If you cannot say how the queue keeps messages durable or how the store picks a leader, the diagram stops counting.

The second reason is level. The prompt is the same for everyone, and the level you are hired at is read off how you handle it. This guide covers who gets the round, how the 45 minutes tend to go, the prompts candidates report, what interviewers appear to reward at L4, L5 and L6, and a three-week plan. Google does not publish a rubric. The specifics come from prep guides and candidate accounts read in November 2026, and the guide flags where they contradict each other.

Who gets the round

Sources agree from L5 upward and disagree below it.

  • L5. A standalone system design round is standard. One guide says one round; another says one or sometimes two.
  • L6 and above. One guide says two rounds. Another describes an optional second round for senior candidates that tests the boundary between L5 and L6.
  • L4. Three sources, three answers. One says a dedicated round "has become standard in 2026". One says it is optional or folded into a coding round. One says design at L3 and L4 usually appears inside a coding round, as a follow-up about scaling the thing you just wrote.
  • L3. No dedicated round in any account.

If you are interviewing at L4, ask the recruiter whether your loop has a design round. If the answer is vague, prepare a lighter version anyway. An hour of design practice is cheap insurance against a coding interviewer who ends with "now suppose this runs on a thousand machines".

For some roles the round is a machine learning design interview instead: a ranking system, a retrieval pipeline, serving infrastructure for a large model. Which one you get is reported to depend on the role and the interviewer's background.

Format, tools and the in-person question

The round is one open-ended prompt and a conversation. Accounts put it at 45 minutes, sometimes up to an hour.

Tools are less settled. One prep site says candidates work in a plain Google Doc or a basic whiteboard with no diagram templates. Another says Google often uses Google Drawings, which does not show your strokes to the other person until you finish a shape, and recommends rehearsing in a tool that is not real-time. Confirm with your recruiter and practise in whatever they name. If nobody can tell you, practise describing your design in words and a text outline, because that works in every tool.

On location, one process guide says the onsite "is usually virtual now, although some roles or locations may bring back in-person final rounds", and two prep sites say some onsites have moved back to in person, partly in response to cheating in remote interviews. If yours is in person, draw on a physical whiteboard at least twice before the day. Handwriting a diagram is slower than you remember.

On AI tools, the reporting is about coding, and it is inconsistent. A May 2026 news article, citing Business Insider, describes a pilot from the second half of 2026 in which junior and mid-level candidates in parts of the United States use Gemini in a code comprehension round. One newsletter says more broadly that Google "will allow AI tools in coding interviews" on a gradual rollout. No account read for this guide mentions an AI assistant in Google's system design round. Assume it is you, the interviewer and a blank page.

The 45 minutes

Nobody hands you an agenda. This split is a suggestion, built from how prep guides describe strong rounds; it is a starting point to adjust, with less time on requirements and more on depth as the level rises.

Minutes What you are doing
0 to 7 Pin down users, the two or three core operations, scale and latency targets
7 to 12 Rough numbers: requests per second, storage per year, read to write ratio
12 to 22 High-level design: components, data flow, the main data model
22 to 38 One or two components in depth, chosen by you
38 to 45 Failure modes, bottlenecks, what you would change at ten times the load

Prompts are reported to be deliberately vague, so the first block is real work. You are expected to define the users and the targets yourself and say them aloud. A candidate who waits to be told the scale has already answered the level question, and not in their favour.

Prompt families candidates report

The lists on prep sites are long. They sort into a few families, and preparing one prompt per family covers more ground than memorizing twenty.

  • Infrastructure building blocks. A managed publish and subscribe service or message queue, a distributed job scheduler, an API rate limiter, a content delivery network.
  • Pipelines. A logs and metrics pipeline with processing in under a minute, a web crawler at Google scale.
  • Google-shaped products. File storage and sync like Drive, video distribution, maps, live chat for a video stream, a real-time global leaderboard.
  • Operations at scale. One reported prompt is simply "upgrade 5,000 servers", which is about rollout, safety and rollback with no product in sight.
  • Machine learning systems. Feed ranking, personalized search for short videos, retrieval-augmented search, serving a large language model.

Infrastructure and operations prompts are where the "what is inside the box" pressure is strongest. If the prompt is a message queue, there is no managed queue to hide behind.

What the rubric appears to reward

There is no public rubric. Prep guides converge on a similar set of things an interviewer writes about.

  • Requirements. Did you scope the problem, and did you scope it yourself.
  • Decomposition. Are the components sensible, and does data flow between them in a way that would work.
  • Depth from first principles. One guide says this is what Google weights most. You are expected to explain or build the component underneath a managed service. Another guide asks for at least one component examined three layers down.
  • Trade-offs with reasons. The alternative you rejected and why, ideally with a number attached.
  • Scale reasoning. Estimates that change a decision. Arithmetic that changes nothing is filler.
  • Production thinking. Failure modes, monitoring, rollout.
  • Communication. Whether the interviewer could follow you and steer you.

Keep in mind where the notes go. The interviewer writes them up for a hiring committee that never saw your diagram. "Chose a log-structured store because writes dominate ten to one, and said what that costs on reads" survives the trip. A beautiful drawing does not.

L4, L5 and L6 on the same prompt

Take a rate limiter. The table is a summary of how guides describe the three bars, applied to one prompt.

Level What a passing answer sounds like
L4 Works within the requirements given. Clean design, a sensible algorithm such as token bucket, a reasonable deep dive when asked. Some guidance is fine.
L5 Raises the important requirements unprompted: per user or per key, what happens across regions, what the caller sees when limited. Picks the hardest component and goes deep without being steered.
L6 Names the constraint that drives the whole design before the interviewer points at it, for example that limits must hold when the counter store is partitioned. Talks about operating it: rollout, monitoring, cost, what breaks first.

The trap for senior candidates is giving the L4 answer well. One guide calls a strong L4-style answer one of the most common routes to a down-level decision at L6. Nothing in it is wrong, and it still does not show the level you asked for.

How it differs from Amazon and Meta

If you are interviewing at more than one of these, the same preparation mostly carries over. Three differences are worth adjusting for.

Amazon. The Amazon system design round is covered in its own guide. Amazon's own preparation page for SDE II says to expect at least one question on software systems design, lists what it looks for as practicality, accuracy, efficiency, reliability, optimization and scalability, and describes 55-minute interviews. Its page says nothing about being pushed below managed services. For Google, one prep site singles that out as the difference from other large companies: naming a managed service is where the questioning starts.

Meta. Meta splits the round in two: system design for infrastructure roles and product architecture for product roles, each reported at 45 minutes. Product architecture spends real time on API shape and the client. None of the Google accounts read for this guide describes a product variant; apart from the machine learning version, you get one style of round, and it leans toward infrastructure. The Meta design interview guide covers both variants.

What happens to the result. At Google your design interviewer does not decide. Their written feedback is one page in a packet read by a committee, so being followed easily counts for more than it does where the interviewer argues your case in person.

A three-week plan

This assumes about an hour a day and that you already know the basic building blocks. If you do not, start with the system design framework and add a week.

Week one: internals. Pick five components you normally treat as a box: a message queue, a key-value store, a load balancer, a cache, a consensus-backed lock service. For each, write one page on how it works inside: how data is laid out, how it replicates, what happens when a node dies. This is the work most candidates skip and the work this round rewards.

Week two: prompts, timed. One prompt a day from the families above, 45 minutes, spoken aloud, in the tool your recruiter named. Record yourself. Listening back, mark every place you named a technology without saying why. Use the questions by level list to check your answer against the bar you are aiming at.

Week three: other people. Two or three mocks with someone who interrupts and asks "how does that work underneath". Self-practice cannot produce that pressure, since you never ask yourself the question you cannot answer. A colleague who has run design interviews will do; if you do not have one, a session with a system design coach who has interviewed at your target level fills the gap. Spend the last two days on your two weakest components, not on new prompts.

Mistakes that cost the round

  • Drawing before scoping. You optimize for a problem the interviewer did not have in mind and find out at minute 25.
  • Managed services as answers. "I would use a managed queue" is fine as a first line and fatal as a last one.
  • A memorized architecture. Interviewers change one constraint to see whether the design moves. If it does not, they know where it came from.
  • Detail in the wrong place. Five minutes on HTTP status codes is five minutes not spent on the part that carries the level.
  • Trade-offs only when asked. Above L4 you are expected to raise them yourself.
  • Running out of time before failure modes. The last block is where L5 and L6 evidence is collected. Protect it.
  • Silence while thinking. The interviewer can only write down what you said.

Where coaching fits

The gap in most people's preparation is the follow-up question from someone who knows the internals. As of November 2026, 12 of the 40 coaches on techinterview.coach list system design; the system design coaches page shows who they are and what a session costs. Two coaches have Google on their history and appear on the Google page. Before paying for anything, you can talk a design through in the practice interview, which has a system design round with an AI interviewer and a written recap afterward. It is free; sign in to start.

Related guides