Why it matters
Interviews increasingly present a working system and ask what you would improve, because it tests judgement rather than construction. There is no single right answer, so the evaluation is on how you find the constraint: reading the symptoms, forming a hypothesis, choosing a measurement and accepting a cost. Candidates who list best practices without connecting them to the described symptom sound like they are reciting a checklist.
Mental model
It is a performance review for an architecture: symptoms, diagnosis, one prescription, follow-up. The metaphor breaks because a patient can describe pain, while a system description omits most of what you would normally measure, so you must ask and the owner may not know either.
How it works
Step 1 of 4
Ask what it optimises
Restate the system's goal and the symptom in one sentence each before naming a component. A cost system and a latency rewrite point in opposite directions.
Ambiguity: low · Time pressure: low
Choose · Make the smallest change
Fix the stage that binds; attach a rollback and the metric that confirms it.
The page's default: the smallest intervention against the constraint, with the new bottleneck named before the change ships.
Axes
- Ambiguity
- one constraint explains it → several candidates
- Time pressure
- no live incident → relief needed now
Figure. Diagnosis and urgency are different axes: a known constraint is fixed, an ambiguous one is measured, and a live incident changes the first move when the candidates are still open, and never settles the diagnosis. The top-right cell is relief, not a fix.
What to notice. The axes are qualitative; no percentile, error rate or budget appears because the page names none. Adding capacity is relief, and the page's caveat is that it fails on a serial dependency. The figure does not claim a rewrite is never right — only that a rewrite under an ambiguous symptom is a guess.
Text version
A two-by-two decision map. The horizontal axis is ambiguity, from one constraint explains it on the left to several candidates on the right. The vertical axis is time pressure, from no live incident at the bottom to relief needed now at the top. Bottom left, one constraint and time: make the smallest change, with a rollback and the metric that confirms it. Bottom right, several candidates and time: measure before changing, because the measurement delays the visible fix and buys a correct diagnosis. Top left, one constraint and a live incident: apply the targeted fix now, because the constraint is known, and attach the rollback. Top right, several candidates and a live incident: relieve first and diagnose after, conditional because capacity buys time without fixing efficiency and will not scale out a serial dependency. Four steps walk the cells in that order, starting with the system's goal.
- Restate the system's goal and the symptom that prompted the review, in one sentence each.
- Ask for the numbers behind the symptom: latency percentile, error rate, cost per request, or human review load.
- List candidate constraints and rank them by how well they explain the symptom.
- Choose one and describe the measurement that would confirm or refute it before changing code.
- Propose the smallest intervention, including the rollback condition.
- State the new bottleneck the fix creates and how you would watch for it.
Trade-offs
| Choice | Good for | Cost | Fails when |
|---|---|---|---|
| Optimise the obvious stage | Fast to implement | May miss the real constraint | The symptom has more than one cause |
| Rewrite the component | Removes accumulated complexity | Large risk and long feedback loop | The constraint is elsewhere in the chain |
| Add capacity | Immediate relief | Cost without fixing efficiency | The bottleneck is a serial dependency |
| Improve measurement first | Correct diagnosis | Delays the visible fix | The incident is live and needs relief now |
Good for
Fast to implement
Cost
May miss the real constraint
Fails when
The symptom has more than one cause
- The symptom has more than one cause
- The constraint is elsewhere in the chain
- The bottleneck is a serial dependency
- The incident is live and needs relief now
Figure. Each choice with what it buys, what it costs, and the case where it loses.
What to notice. The fourth column is the decision: if the losing case describes your situation, the choice is wrong.
Text version
Four trade-offs: Optimise the obvious stage is good for Fast to implement, costs May miss the real constraint, and fails when The symptom has more than one cause. Rewrite the component is good for Removes accumulated complexity, costs Large risk and long feedback loop, and fails when The constraint is elsewhere in the chain. Add capacity is good for Immediate relief, costs Cost without fixing efficiency, and fails when The bottleneck is a serial dependency. Improve measurement first is good for Correct diagnosis, costs Delays the visible fix, and fails when The incident is live and needs relief now
Failure modes
- Checklist answer. The response lists caching, batching and observability without tying them to the symptom. Debug: check whether each proposal names the measurement it would move.
- Ignoring the goal. The system is optimised for cost while you propose a latency rewrite. Debug: ask what the team is measured on before ranking fixes.
- No cost side. Every proposal is an improvement with no downside, which means the trade-off was not found. Debug: ask what the change makes worse and see whether the answer is specific.
In production
Rehearse on systems you have actually operated, and bring one number per story. A drill answer without a measurement is an opinion, and interviewers score it as one.
Common misconceptions
People think: There is a correct answer to an improve-this-system question.
Actually: The evaluation is the diagnostic process and the trade-off, not the specific component you pick.
People think: More instrumentation is always the best first move.
Actually: It is the right move when the symptom is ambiguous; when the constraint is obvious, delay costs credibility.
Interview angle
Strong answers ask for the goal and the numbers, rank hypotheses, pick one measurement and commit to a small change with a rollback. They name the new bottleneck. Weak answers list generic improvements and never connect them to the symptom.
Sources
- Will Larson, Staff Engineer guides — framing technical direction with evidence.
- Google SRE Book — objectives and error budgets as the constraint.
- Kleppmann, Designing Data-Intensive Applications — reasoning about which component binds first.