Phased Delivery & the Interview Clock
Budget the 45 minutes across six phases with exit criteria, reach a complete simple design before adding complexity, and narrate every transition.
A system-design round is a time-boxed delivery problem
You have roughly 45 minutes, one interviewer, and one goal: end with a working design plus enough depth to prove you can build it. The single biggest reason strong engineers fail this round is not weak knowledge, it is pacing. They spend 15 minutes perfecting requirements, draw half a diagram, and the timer ends before there is anything to deep-dive on. A repeatable phase structure with an explicit minute budget prevents that.
The canonical 6-phase clock for a 45-minute round
Phase Budget Exit criterion
1 Clarify + scope + NFRs ~5 min You have functional + non-functional reqs and the constraint that matters
2 Estimation (entities/QPS) ~2 min You have a read/write QPS and rough storage number to size with
3 API surface ~5 min The 3-5 core endpoints (or events) are named with inputs/outputs
4 High-level design ~15 min A complete boxes-and-arrows design where every functional req is satisfied
5 Deep dive(s) ~10 min The tightest NFR bottleneck is addressed with a committed choice
6 Wrap-up ~3 min Top remaining bottleneck, failure mode, monitoring, cost driver stated
Notice that phases 1 and 2 together take only about 5 to 7 minutes. Requirements and estimation are the setup, not the main event. The bulk of the clock, phases 4 and 5, goes to design and depth, because that is what the interviewer is actually scoring.
The prime directive
Reach a COMPLETE working design before you add any complexity. A simple design that satisfies every functional requirement beats an elaborate half-design every time. Do not shard, add Kafka, or optimize the cache until the plain version works end to end.
Two skills make the clock work in practice. First, exit criteria. Each phase has a concrete condition that tells you it is done and you may move on. Without one you drift. When you have a read QPS and a storage estimate, estimation is over, stop refining the number. Second, narrated transitions. You say the phase change out loud so the interviewer follows your lead: "I have a working design now, so let me harden the availability, which is the tightest requirement here." This keeps you visibly in control and signals seniority.
Interview nuance: treat the framework as a scaffold, not a script. If the interviewer jumps you to the data model in minute 3, follow them, then loop back to fill the gaps. Reordering on their cue is a strength. Rigidly reciting phases while they try to steer is the tell of a memorized answer.
Interview nuance: if you are running long, say so and cut. "I am watching the clock, so I will lock the high-level design and move straight to the delivery bottleneck." Interviewers reward candidates who self-correct pacing over ones who need rescuing.
Recap: budget about 5 to 7 minutes for requirements and estimation, spend the bulk on design and deep dives, use an exit criterion to leave each phase, and narrate every transition so you visibly lead the round to a complete design.
Apply
Your turn
The task this lesson builds to.
Produce a labeled 6-phase walkthrough plan for a 45-minute 'Design a URL shortener' round, with a minute budget per phase and the exit criterion for each.
Think about
- How much time goes to requirements+estimation vs design+deep dives?
- What is the exit criterion that lets you move to the next phase?
- How do you narrate transitions so the interviewer follows your lead?
Practice
Make it stick
A second problem on the same idea, so it survives past today.
Produce a labeled phase plan for a compressed 35-minute 'Design Twitter/X home timeline' round where the interviewer has told you up front they care most about read fanout at 500M daily active users. Show how you re-budget the shortened clock and where you cut.
Think about
- When the interviewer pre-declares the scoring focus, which phases shrink and which grow?
- What is the minimum breadth you still need before the deep dive is credible?
- How do you quantify why fanout is the thing that matters at 500M DAU?