The Kafka Bill — Week 3 CheckWeek 3 of Kafka in Production: cost physics. Cross-AZ geometry, follower fetching, tiered storage, diskless topics, and the sizing math — all from this week's case studies.

RF=3 across 3 AZs, leaders evenly spread, leader-only fetching. What fraction of consumer fetches cross an AZ boundary?

What did enabling rack-aware follower fetching cost Grab in exchange for the ~25% cross-AZ traffic reduction?

In KIP-405 tiered storage, which segments are eligible for upload to the remote tier?

Your remote object store has a multi-hour outage. What happens to local broker disks on a tiered topic?

Which cost does KIP-1150 (Diskless Topics) eliminate that KIP-405 tiering never touches?

Per AWS's sizing doctrine, sustained cluster throughput is the minimum of three bounds. Which set is correct?

Why do burst credits make undersized Kafka clusters especially dangerous?

A backfill consumer starts reading 3-day-old data on a busy cluster and tip consumers' latency degrades. What's the mechanism?