The mistake that costs you the hour
You will not fail on knowledge. You will fail by reaching for the marker, the way the first candidate did.
That costs more than it looks. Load balancer, three app servers, a database, ninety seconds, and every one of those boxes is an answer to a question nobody asked. It looks decisive, and it is the strongest negative signal in the format.
What the drawing silently claims
The diagram makes claims nobody agreed to. Boxes drawn before requirements assert a scale, a consistency model and a read-write ratio nobody agreed to. When the interviewer asks why there is a cache, the only answer available is that caches are normally there. Every question after that lands on sand.
What gets scored in the opening is narrower than you think. Do you notice the prompt is underspecified. Do you ask about scale before choosing components. Do you cut scope on purpose. Do you commit to numbers you can be held to. Those four beat any amount of component knowledge, because they are what the job actually looks like. Real design documents start from constraints.
How to spend the ten minutes
It is a conversation, not a checklist. Ask two or three questions, propose your own answer when the interviewer deflects, and state the scope back before you move on. They deflect deliberately, and “you decide” is an invitation to show judgement rather than a refusal. Say “I will assume 10 million daily actives, read-heavy at about 100 to 1, tell me if that is wrong” and you have shown more than six questions would.
Budget about ten minutes of a 45 minute interview. Long enough to fix the constraints, short enough to leave room for the design they imply.
Worked example
Two candidates get "design a URL shortener" in the same week. The first starts drawing immediately: client, load balancer, two app servers, Postgres, Redis. At minute 20 the interviewer asks why Redis is there, and the honest answer is that shorteners usually have a cache. At minute 30 they are asked about the ID scheme and pick auto-increment, which turns out to leak volume to competitors and to break the multi-region requirement nobody established. The debrief says "knows the components, could not justify any of them."
The second spends nine minutes asking. How many URLs created per day, roughly 100 million. Read to write ratio, they suggest 100 to 1 and the interviewer nods. Custom aliases, out of scope. Analytics, out of scope. Link expiry, yes. They say back: "so 100M writes a day, 10B reads, short links, no custom aliases, expiry supported, and I care most about read latency." Every later decision refers to those numbers, and when they choose base62 over auto-increment they justify it in one sentence. Same knowledge, same components, opposite outcomes.