Section 3: The Arithmetic That Ends Arguments - Scale Estimation & Workload Analysis

Executive Summary: This section builds the numbers ledger that every later section will spend. From one stipulated figure — 200M daily active users — we derive the read load, the write load, the 10:1 ratio between them, the long-tail distribution that makes the median post tiny, the hot-row figure that makes one celebrity post dangerous, and the storage and cache footprints. The centerpiece is a single comparison: a hot post's write rate on one row exceeds the entire platform's peak write rate — proof that the hard problem here is local contention, not global throughput. Get this arithmetic right and the architecture in Section 6 practically draws itself.

The interviewer has stipulated the scale: a mid-size social product, 200M DAU. Weak candidates nod at that number and move on. You are going to spend it — because estimation in a Staff interview is not a ritual of multiplication, it is the phase where you discover which problems are real. Both failure archetypes from Section 1 die here if you do the arithmetic honestly: the Dismissal never computes what one hot row endures, and the Citadel never computes how small the median post actually is.

The Read Side: Every Rendered Post Is a Count Read

Start where the load actually lives. A like count is not read when someone cares about likes — it is read every single time a post appears on a screen, because the count (and the viewer's own heart state) is part of the post's rendering. When Priya scrolls her feed on the subway, every post that slides past fires a count read whether she glances at it or not.

Assume an engaged user views 50 posts per day — a few feed sessions, a dozen posts each. Then:

• 200M DAU × 50 posts viewed/day = 10B post views/day
• 10B ÷ 86,400 seconds/day ≈ ~116K count reads/s average
• Traffic is diurnal, not flat — evenings in each major region stack into peaks roughly 3× the daily average. ×3 ⇒ ~350K reads/s peak

Why ×3 and not ×10? Because a global social product's regional peaks partially offset each other; a single-region product or a product with one dominant time zone would deserve a harsher factor. Saying why your peak factor is 3 is worth more than the 3 itself.