Section 9: The Escalation Gauntlet - Interviewer Deep-Dive Challenges

Executive Summary: The interviewer stops asking how the design works and starts breaking it. Six escalations follow: a 5M-follower publish during Maya's 5,000 views/s spike, a 20-minute CDN purge outage, the loss of the primary region, a bot army of 100K accounts clapping, a 300,000-word paste, and "why not just WordPress?". Each answer is a mechanism, a number and an honest residual risk, and every one of them leans on a decision already made in Sections 5 through 8.

It's 10:20 on Tuesday. Maya's post is at its peak and her typo fix has just gone out. The interviewer smiles: "Good. Now let's make your day worse."

You: "Go ahead. For each, I'll say what keeps working, what degrades, and what pages someone."

Challenge 1: The Giant Publishes Mid-Spike

Interviewer: "A publication with 5 million followers publishes at the exact moment Maya's post is doing 5,000 views/s."

You: "Neither system feels much. The giant is far above the 100K-follower threshold, so it's on the pull path: publishing writes one revision row, one pointer flip and one outbox row. No fan-out burst. Pure push would land 5M inbox inserts at once, about six hours of our normal ~230/s average fan-out in a single spike. Instead, each of those followers' feed loads merges one more cached recent-posts list on top of the ~3 giants a typical reader already follows, and the ~1,750/s feed peak barely notices.

"Maya's article is a different URL, served from the edge at close to a 100% hit ratio. The one coupling I'd check is the outbox: the purge worker and the fan-out worker are separate Kafka consumer groups, so Maya's 10:20 purge never queues behind anyone's fan-out."

The flip trigger comes from Section 7's hybrid feed: if push-path fan-out lag p99 approaches the 5-minute SLO, lower the threshold so more publications move to pull.

Quick check before moving on: