Section 13: The Judges' Table - Final Summary & Candidate Scorecard

Executive Summary: This closing section turns the module into a way to grade yourself. We replay the arc as signal, ledger the eleven decisions with their flip triggers, lay out a 100-point scorecard in Kafka terms with score bands, and list the prompts that are this question in disguise.

The panel never graded the box labeled "broker." They listened for whether you heard "dozens of teams, in order, replaying a week of history" and answered "a log, not a queue." And when broker 12 died with offsets up to 1,284,553,905 on its disk and the last committed offset at …903 (high watermark …904), they listened for "…904 and …905 were never acknowledged, broker 3 leads from the ISR, and Ari's retry is deduplicated by sequence number" versus "Kafka replicates everything, so nothing is lost."

The Arc, Replayed as Signal

1. Framing (Sections 1–2): per-message acks, global order and delete-on-consume each died in a clarification.
2. Estimation (Section 3): 1.5 M messages/s at peak; 907 TB on disk sets 30 brokers at 63% full, while egress runs at 11% of the NIC.
3. API and data (Sections 4–5): produce returns (partition 37, offset 1,284,553,901); segments named by base offset with a sparse index.
4. Architecture (Section 6): 200 leaders per broker, three KRaft controllers, one replica per zone, reads up to the high watermark.
5. Deep dives (Sections 7–9): about 11 minutes of page-cache tail; acks=all with min.insync.replicas 2; frd-18's output and offset in one transaction.
6. Failure and growth (Sections 10–12): about 10 s without a leader for 1/30 of the cluster; zone B dark, with about 10 s leaderless on a third of the partitions and writes then continuing at ISR 2; earlier decisions flipping into tiered storage and share groups.

Key Insight: None of those signals is "knows Kafka's configs." The panel tests which mechanism keeps which promise: the partition keeps per-key order, the ISR and high watermark keep an acknowledged write, sequence numbers keep a retry from duplicating, and the transaction keeps an output with its offset. None of them reaches past Kafka into Postgres.