Table Operations — Week 2 CheckWeek 2 of Lakehouse & Table Formats: the maintenance loop — compaction, snapshots, schema changes — from Ancestry's 100-billion-row table to Kakao's 3-billion-a-day logs.

Mechanically, why do small files degrade queries even when total data volume is unchanged?

Miles Cole's benchmark compared five compaction postures on trickle ingestion. The headline result:

Ancestry keeps hourly MERGEs affordable on a 100B-row table with merge-on-read. What did that choice OBLIGATE them to do?

Storage-partitioned joins removed the dominant shuffle cost inside Ancestry's MERGE. What are the preconditions?

Kakao runs UPSERT for CDC-fed DB tables but APPEND for 3B-a-day server logs. Why the split?

Kakao's KST hour-partitions landed 9 hours off. The mechanism?

Kakao dropped Flink checkpoints from 3 minutes to 1 minute. What did they buy and what did they pay?

Whoop's Glacierbase applies Flyway-style discipline (versioned, hash-validated, reviewed migrations) to Iceberg DDL. The strongest justification: