Running a system design interview

The poster maps what a system design interview grades. This article follows one interview minute by minute from both sides of the table: what the interviewer is doing in each phase, what the candidate is expected to produce, what strong and weak answers look like, and how the notes become a decision.

If you run interviews, read it as a script and a scoring guide. If you are the candidate, it shows what the person across the table is listening for. The time plan, the scoring example and the noise figures are computed by code, and the calculator fits the plan to your slot.

Two sides of one table

In a system design interview you draw a diagram, but the conversation is what gets graded. The interviewer is not waiting for a correct diagram, because there is none. Their job is to collect evidence on five signals in a fixed time, and write it down. The candidate’s job is to make that evidence easy to see.

  1. Clarifying requirements. You ask what to build, for whom and at what scale before drawing a single box.
  2. Trade-offs. You spell out the price of every choice: what you gain and what you give up.
  3. Estimation. You rough out the scale in orders of magnitude and let those numbers shape the design.
  4. Depth. You dig into one or two components properly instead of name-dropping ten.
  5. Communication. You think aloud, pick up on hints and drive the discussion instead of waiting to be prompted.

All five on one page: the poster →

The rest of the article follows from that. Phases give each signal a moment to show. Hints let a stuck candidate still show the next one. Scoring makes the decision rest on what was observed, not on how the hour felt.

1. Before the call

The interviewer

  • Use a question you have solved yourself, ideally from a family you know well (the poster lists nine). You need to recognise a good answer that is not yours.
  • Prepare three deep-dive branches and the follow-up constraint for each that separates a senior answer from a mid-level one. For a URL shortener: ID generation, the redirect cache, analytics.
  • Write down the bar and the decision rule now, for the level you are hiring for, before you meet anyone. Section 6 shows why.
  • Plan the clock. Know at what minute the deep dive must start (the calculator below gives it).

The candidate

  • Practise aloud and against a timer: the format punishes silence more than mistakes.
  • Know the numbers on the poster well enough to round them in your head.
  • Ask at the start how the interviewer likes to work: shared doc or whiteboard, interrupt freely or hold questions.

The first two minutes

The interviewer sets the frame, briefly: “The question is vague on purpose. Ask me anything, think aloud, and expect me to interrupt: that is me steering, not a bad sign.” A candidate who hears this relaxes, and a relaxed candidate gives more evidence.

2. The clock

The poster’s 45 minutes are design time, not the length of the meeting. A calendar hour also holds an introduction and the candidate’s own questions, and a 45-minute slot leaves noticeably less than 45 minutes to design.

Fit the plan to your slot

The poster’s 45 minutes are design time. Enter the length of the call and what you keep for the introduction and the candidate’s questions; the phases keep the poster’s proportions.

PhaseMinClockShare of design
Introduction50:00–0:05
1. Requirements50:05–0:1011%
2. Estimation50:10–0:1511%
3. API and data model70:15–0:2216%
4. High-level design100:22–0:3222%
5. Deep dive130:32–0:4529%
6. Bottlenecks and failures50:45–0:5011%
Candidate’s questions100:50–1:00

45 of 60 minutes go to design. The deep dive gets 13 minutes and should start by 0:32; wrap-up starts at 0:45.

Minutes are split in proportion to the poster’s plan and rounded so they add up exactly (largest remainder). The clock counts from the start of the call.

The interviewer owns the clock. In the default hour the deep dive gets 13 minutes and must start by 0:32. If at that moment the candidate is still adding boxes, interrupt and pick the deep dive yourself: losing polish on the high-level picture costs little, losing the deep dive costs the most informative signal of the interview.

When the slot is short, compress the API and the data model, not the requirements. Two minutes saved on requirements usually cost ten in redesign.

3. Phase by phase

Times are for the default hour. For each phase: what the interviewer is doing, what the candidate should leave on the board, what separates a strong answer from a weak one, and one question that reliably produces evidence.

1. Requirements 0:05–0:10, 5 min

Core features, non-functional targets (latency, availability, consistency) and what you are leaving out.

InterviewerReads the prompt, then mostly listens. Answers with numbers when asked (“100 million new links a month”) and says “your call” when the choice belongs to the candidate.
Candidate producesA short written list: functional scope, non-functional targets, what is out of scope, the one or two numbers that matter.
Strong signalAsks the questions that change the design (read/write ratio, consistency, scale), writes the answers down and stops.
Weak signalStarts drawing in the first minute, or asks twenty questions without saying which answers matter.
A probe to ask“Which of these requirements would change your design the most?”

2. Estimation 0:10–0:15, 5 min

Read and write QPS, storage and bandwidth — the load your design has to survive.

InterviewerLets the arithmetic run. Corrects an order of magnitude only if it would derail everything after it.
Candidate producesRead and write QPS, the peak, storage and bandwidth, each with one line of working.
Strong signalRounds boldly and says what each number implies: “two thousand writes a second, one database copes”.
Weak signalPrecise arithmetic that leads nowhere, or skipping it with “it’s big”.
A probe to ask“What does this number tell you about the design?”

3. API and data model 0:15–0:22, 7 min

A few endpoints and the main entities, tied to the access patterns they need to support.

InterviewerChecks that the endpoints serve the requirements agreed earlier and that every entity has a reason to exist.
Candidate producesThree to five endpoints with request and response; the core entities with their keys and access patterns.
Strong signalPicks the store from the access pattern and names the key the data will be partitioned on.
Weak signalFull CRUD for every entity; a database chosen by brand.
A probe to ask“Which query runs most often, and what serves it?”

4. High-level design 0:22–0:32, 10 min

Client, balancer, services, cache, database, queue — then trace a real request end to end.

InterviewerAsks for one request to be walked through end to end and notes components that have no reason to be there.
Candidate producesThe box diagram, with a read and a write traced through it.
Strong signalEvery box is justified by a requirement or by a number from the estimate.
Weak signalThe standard picture drawn from memory, with a queue and a cache nobody needs.
A probe to ask“Walk me through what happens when a user does this.”

5. Deep dive 0:32–0:45, 13 min

The trickiest piece — the interviewer often picks it: schema, sharding, caching or consistency.

InterviewerIf the candidate has not picked a component, picks one, usually the riskiest. Keeps adding constraints until the candidate reaches the edge of what they know.
Candidate producesDetail on one or two parts: schema, partitioning, consistency, the algorithm, failure handling.
Strong signalCompares two or three options, picks one and states its price; adapts when a new constraint arrives.
Weak signalStays at box level, or defends the first idea against every new fact.
A probe to ask“Traffic grows tenfold. This node dies. Two users do this at the same moment. What changes?”

6. Bottlenecks and failures 0:45–0:50, 5 min

What breaks first, hot keys, single points of failure, monitoring and what you would tackle next.

InterviewerLeaves room for the candidate to criticise their own design, then asks what they would build next.
Candidate producesWhat breaks first, the single points of failure, what to monitor, the next step.
Strong signalFinds the weak spot before being told and ranks the fixes by impact.
Weak signal“It scales horizontally”, with no detail and no metrics.
A probe to ask“What would you monitor, and which alert wakes someone up?”

4. Hints and interruptions

A hint helps you get the next piece of evidence. Use the lightest one that works and note which one you gave: the record should say what the candidate did with it.

StepWhat the interviewer saysWhenHow to note it
WaitNothing, for ten to twenty seconds.The candidate is thinking productively.Not a hint. Silence is part of the format.
Open question“What else could go wrong here?”Something important is missing.Whether they found the gap themselves.
Pointed question“What if two users buy the last seat at once?”The gap is specific and the clock is running.The gap, and how quickly they closed it.
Direct hint“Consider a conditional update instead of a lock.”The candidate is stuck and the phase is overrunning.The hint, then grade what they build on it.
Take over“Let’s assume X and move on to Y.”A phase threatens to eat the deep dive.Not a failure by itself: you are protecting the signal.

Two rules keep this fair. Give every candidate the same ladder, in the same order. And grade the response to the hint, not the fact that one was needed: a candidate who takes a pointed question and builds a sound design on it has still shown the skill you are grading.

5. Same question, different bar

The question rarely changes with the level. What changes is how much of the interview the candidate drives and how far the deep dive goes before it hits the edge of their knowledge.

DimensionMid-levelSeniorStaff
Who drivesFollows the structure, with prompts.Drives the whole interview; the interviewer mostly reacts.Drives, and reframes the problem when the prompt asks the wrong question.
RequirementsA functional list and a few numbers.Non-functional targets that visibly shape the design.Business constraints, cost, operations, and what not to build.
DepthOne component in detail, with help.Two components, alternatives compared unprompted.The hardest part solved, plus how it fails and how it evolves.
Trade-offsNames them when asked.States the price of every choice.Weighs them against the organisation: team, migration, cost.
FailureHandles a node going down.Hot keys, retries, consistency during a partition.Blast radius, degraded modes, rollout and rollback.

Decide the level you are interviewing for before the call. Grading “strong for a mid-level” and then proposing them as senior is how levels drift.

6. From notes to a decision

Score each of the five signals separately, after the call, against fixed anchors:

ScoreMeaningTypical evidence
4 · StrongAbove the bar for the level.Raised the issue before being asked; compared options with numbers.
3 · MeetsAt the bar, minor gaps.Got there with an open question at most.
2 · MixedBelow the bar on part of the signal.Needed pointed questions; the right idea, shallow reasoning.
1 · MissingNo evidence, or wrong.Needed a direct hint and could not build on it.

Then apply a rule you wrote down before the interview. A plain average is tempting, but it hides what matters:

Same average, different decisions

Signals 1–5AverageBy averageBy the rule
Even333333.0hirehire
Brilliant with a hole444123.0hirereject
Just below323332.8rejectreject
Scores 1–4 per signal. The rule, fixed before the interview: average at least 3 and no signal below 2. Signals: 1 Clarifying requirements · 2 Trade-offs · 3 Estimation · 4 Depth · 5 Communication.

The first two candidates both average 3.0. By average alone, both pass. The rule (average at least 3, nothing below 2) stops the second, whose depth scored 1: a candidate who cannot go deep on any component will struggle to own one. The third misses the average by a little and has no hole; that case is worth discussing at the debrief rather than deciding automatically.

7. Several interviewers, independent notes

One interviewer’s score is the candidate’s level plus noise: the question they happened to draw, the interviewer’s mood, the first impression. Loops have several interviewers to average that noise away, and the arithmetic says when that works.

Noise left after averaging the panel

How the interviewers judged1 interviewer246
Independently, wrote before the debriefcorrelation 0100%71%worth 2.050%worth 4.041%worth 6.0
Somewhat alignedcorrelation 0.3100%81%worth 1.569%worth 2.165%worth 2.4
Talked first, then wrotecorrelation 0.6100%89%worth 1.384%worth 1.482%worth 1.5
Error of the panel’s mean score as a share of one interviewer’s error, √(ρ + (1 − ρ)/k). “Worth n” is how many independent interviewers would give the same precision.

Four interviewers who judge independently cut the noise to 50% of one interviewer’s. Four who talked before writing, with errors correlated at 0.6, keep 84%: the panel is worth about 1.4 independent interviewers, and no number of extra people takes it below 77%. Hence the practice: everyone writes and submits feedback before the debrief, and nobody reads anyone else’s first.

A separate line of research explains why the format is structured at all. Meta-analyses of hiring methods find that structured interviews (same questions, fixed anchors, scores recorded per dimension) predict job performance better than free-form ones.

8. Writing the feedback

Write it the same day, per signal, with evidence rather than adjectives. “Smart” and “good communicator” cannot be checked; a quoted decision with a timestamp can.

Recommendation: hire, senior
Question: URL shortener (deep dive: ID generation)

Requirements (3)  asked read/write ratio and retention unprompted;
                  missed analytics until 0:12.
Trade-offs (4)    "a counter range per server: no coordination per
                  request, gaps in IDs are acceptable".
Estimation (3)    write QPS and storage right within 2x;
                  did not derive cache size.
Depth (2)         two ID schemes compared, collision rate argued;
                  needed a pointed question on range exhaustion (0:38),
                  then proposed fetching a new range, missed the stall-free swap.
Communication (4) drove the pace, checked in at every phase boundary,
                  stated assumptions before choosing.

Hints: 1 pointed question (0:38).

Someone who never met the candidate should be able to reach the same scores from this note. If they could not, the note is a verdict, not feedback.

9. Mistakes on both sides

MistakeWhy it hurtsInstead
Waiting for “your” designThe candidate is graded on matching a hidden answer, not on reasoning.Grade the reasoning. Any coherent design with stated costs can score a 4.
The silent interviewerThe candidate cannot tell whether they are on track, and you collect less evidence.Short acknowledgements; steer at phase boundaries.
Trivia gotchasThey measure memory, not design.Ask how they would find out, or give the fact and watch what they do with it.
Grading the final diagramThe diagram only supports the conversation. The evidence is in the decisions and their reasons.Note decisions and reasons as they happen, with timestamps.
Deciding in the first five minutesThe rest of the interview only confirms the first impression.Score each signal separately, after the call, against the rubric.
No time left for the deep diveThe most informative phase is never observed.Use the checkpoints from the clock and take over when a phase overruns.
Candidate: treating it as an examThey wait to be told what to do, so “who drives” scores low.Check in at phase boundaries: “shall I go deeper here or move on?”

The candidate-side list on the poster (buzzword stacking, ignoring failure, fighting hints, breadth without depth) mirrors this table: each one hides the evidence the interviewer is looking for.

A one-minute checklist

Interviewer

  • Question solved, three deep-dive branches ready, bar and rule written down.
  • Deep dive starts by 0:32; take over if a phase overruns.
  • Lightest hint first, every hint noted.
  • Scores per signal, after the call, before reading anyone else’s.

Candidate

  • Requirements first, with numbers, written down.
  • Let every box earn its place from a requirement or an estimate.
  • Say the price of each choice; welcome the interruptions.
  • Criticise your own design before the interviewer does.

Where this comes from