Deep Dive Selection

The Time Allocation Problem

A baseline design plus estimates consumes ~20 of 45 minutes. The remaining time goes deep on one or two components — and which ones you pick is itself a graded decision. Depth in the wrong place (auth flows for a feed design) wastes the interview’s scoring window.

The Selection Criteria

Rank candidate components by three questions:

 1. RISK        where does this system most plausibly fail at scale?
                (the bottleneck analysis already answered this)

 2. DIFFERENTIATION   what makes THIS system not-generic?
                (URL shortener → id generation + redirects at scale;
                 ride-share → geo-matching; chat → fan-out + ordering)

 3. INTEREST SIGNAL   what does the interviewer keep circling back to?
                (repeated follow-ups = invitation to go deeper there)

Score each component mentally; dive where all three align.

RideShare Selection Walk

 candidates:
 auth/PSP integration    commodity — everyone knows it          score low
 trip CRUD               standard relational patterns           low
 location pipeline       100k writes/sec, ephemeral, sharded    HIGH (risk)
 matching engine         sub-second, geo-indexed, business-core HIGH (all three)
 history/feed            read replicas + cache, well-trodden     medium

 CHOSEN DEEP DIVES: matching engine, then location store
 announced explicitly: "the interesting problems are matching
 and location scale — I'll spend us there unless you'd rather
 I detail something else"

The announcement matters: it hands the interviewer a steering wheel while demonstrating prioritization.

What a Good Deep Dive Contains

Depth is structured, not rambling:

Layer	Content
Problem restated numerically	”600 matches/sec against moving driver positions”
Core mechanism	Geo-index choice with alternatives rejected briefly
Hard edge cases	Contention when one driver gets multiple offers
Failure behavior	Matcher down → queue requests, degrade ETA display
Scale evolution	What doubles at 10x cities

Each layer is one to two minutes. Numbers from earlier estimation get consumed here — that’s why they were computed.

Steering Dynamics

Deep dives are conversations with direction:

 interviewer hint                    correct response
 "how does matching actually work?"  → the invited dive begins, structured
 "what about surge pricing?"         → either scope-park or trade ("that's a
                                       real subsystem — worth trading out the
                                       location internals discussion?")
 silence after your dive             → offer the second-ranked candidate

Candidates who let interruptions scatter them across five shallow topics lose the depth signal entirely; the park-and-trade reflex protects it.

Interview Framing

Interviewers consistently report the same failure mode: complete-looking diagrams, zero depth anywhere. The fix is deliberate allocation — say aloud what deserves the minutes and why, tie it to estimated numbers and stated risks, then execute one structured dive. Choosing depth is designing.