Formats & Internals — Week 1 CheckWeek 1 of Lakehouse & Table Formats: the commit machinery under Iceberg, Delta, and Hudi. Every question traces to a case study or analysis you read this week.

Object storage has no atomic multi-file operation. How do table formats deliver atomic commits anyway?

Vanlightly's model-checking found that multi-writer Delta on plain S3 (historically no conditional PUT) without an external lock provider…

Two Delta writers race for log version N. The loser doesn't always abort — what decides retry vs abort?

Hudi's conflict detection is at FILE GROUP granularity (finer than Delta's partitions). What's the payoff and the catch?

Why did Adobe say Iceberg structurally eliminated their partial-write corruption bugs (~1M batches/day)?

Merge-on-read vs copy-on-write in one sentence: what are you actually choosing?

The Onehouse format comparison is technically substantive AND vendor-authored. What's the professional way to read it?

Delta writes a Parquet checkpoint every ~10 commits. What problem does that solve?