</> MAANG.io

Expert

Design landmark systems in depth, from Dynamo, GFS, Bigtable and Kafka to payments and ads pipelines.

0 of 232 solved 0%

Section 12: Pull Every Lever - Interviewer Challenges & Follow-up Evolution

Executive Summary: The interviewer now tries to break the design six ways, then asks how it grows. Each answer follows one pattern: put a number on the threat, name the lever Sections 7 to 11 already built, and say when a decision from the ledger flips. The evolution half adds five features as new stages or candidates behind the same front door.

The interviewer leans back. "It survives a region loss. Now let me push."

Six stress tests paired with their levers: 10x spike shrinks the cascade, 30 ms budget fuses and prefetches, broad targeting flips D3 to filtered ANN, $200 overspend flips D9 to spend reservations, lost identifiers bring contextual and modeled signals, model regression widens online guardrails

Challenge 1: The Super Bowl Spike

"10x traffic for 30 minutes. Go."

What they're really asking: do you know which resource you cannot buy for an evening?

Ten times the 1.2M QPS peak is 12M QPS. At 500 heavy scores per request that is 6B scores/s, or 40,000 accelerators at 150K scores/s each. We own 4,800. Nobody acquires 35,000 accelerators for half an hour.

The game is on the calendar and the spike lands mostly in the US regions, so the CPU tiers (front door, leaf replicas, light ranker) pre-scale days ahead to those regions' share, not 10× globally. The front door sheds low-value placements first (stories, lower slots), also capping retrieval. Accelerators are fixed, so the cascade bends well past the ladder's 300-candidate cut: 500 candidates for the ~5% highest-value requests, ~25 for the rest, an average of 0.05 × 500 + 0.95 × 25 ≈ 49. Then 12M × ~50 = 600M scores/s, exactly what today's fleet serves. EMPTY stays the last rung of the ladder (D7), never a feed error.

The senior line: "The cascade width is my throttle, not the fleet size."

Quick check before moving on:


Sign in to continue reading

The rest of this lesson is available with a free account. Signing in with Google or Microsoft is free.

Sign in to read the full lesson

Sign in to MAANG.io

Use your Google or Microsoft account — no password to remember.

Continue with Google Continue with Microsoft

Please accept the terms above to continue.