The lakehouse mental model — lake vs warehouse vs table formats vs zones

At a glance

Learning objectives

• Contrast lake, warehouse, and lakehouse by what each does to raw files
• Name the table format's jobs — snapshots, manifests, hidden partitioning, schema evolution
• Place the medallion zones over a lakehouse and say who owns each zone

What you need

• An object-storage lake or lakehouse you can point at
• The Iceberg spec and Delta docs — both are in the DE//AGENT registry
• No installs — this lesson is the map, not the tools

Steps

1. Start from the gap: object storage gives you a place to put files but no structure for trust or ownership. (from pattern-medallion-zones )
2. Name the setting: a lake or lakehouse where many teams write and read, and consumers need to know how far to trust a table. (from pattern-medallion-zones )
3. The table format is what turns the lake into a lakehouse: Iceberg's spec covers the table format, snapshots and manifests, hidden partitioning, schema and partition evolution, and maintenance procedures like compaction and snapshot expiry. (from ref-iceberg-docs )
4. Delta is the other major format: its official docs cover ACID transactions on object storage, time travel, OPTIMIZE and Z-ORDER clustering — and the registry says compare with a Delta vs Iceberg decision guide before committing to a format. (from ref-delta-docs )
5. Multi-format is normal, not a compromise: Amazon EMR's multi-catalog queries Iceberg, Delta Lake, Hudi, and Hive tables through a single catalog and joins data across AWS accounts without copying it. (from ref-query-across-accounts-and-table-formats-with-multi-catalog-in-am )
6. Give the house its zones: raw/ is immutable, exactly as received, never edited, schema-on-read; curated/ is cleaned, typed, deduplicated, quality-gated; serving/ is modelled marts optimised for consumption, SLA-bearing — and promotion between zones is a pipeline step with a gate, not a copy. (from pattern-medallion-zones )

Check yourself

• You can say what object storage lacks — structure for trust or ownership — and which zone supplies it for a given consumer.
• [object Object]
• You can explain why promotion between zones is a gated pipeline step, not a copy.

Worked example

raw/       — immutable, exactly as received; never edited; schema-on-read.
curated/   — cleaned, typed, deduplicated; partitioned; quality-gated.
serving/   — modelled marts optimised for consumption; SLA-bearing.
Each zone has a different retention, quality bar and owner. Promotion
between zones is a pipeline step with a gate, not a copy.

Query Iceberg, Delta Lake, Hudi, and Hive tables through a single
catalog and join data across AWS accounts without copying it.

Sourced from

• Medallion Zones (raw / curated / serving)
• Apache Iceberg Documentation
• Delta Lake Documentation
• Query across accounts and table formats with multi-catalog in Amazon EMR

Security scan

Last sweep 2026-10-24: 0 finding(s), 0 marked as examples.

Common gotchas

• Object storage alone gives no structure for trust or ownership — zones and formats are what turn a lake into a lakehouse.
• The table format is a specification; the spec is the ground truth when tools disagree.
• Do not let curated become a dumping ground; enforce the gate.

Production considerations

• The spec is essential reading for multi-engine shops where Spark, Trino, and warehouses share tables — start there, not at engine docs.
• Cross-account joins without copying are possible through multi-catalog; every extra copy is extra storage and another pipeline to run.

Diagram

lakehouse mental model
   |
   |- object storage --> files, no structure for trust or ownership
   |
   |- table format (Iceberg | Delta) --> ACID, snapshots, manifests, evolution
   |
   |- raw (immutable) --> curated (quality gate) --> serving (SLA)
   |
   |- multi-catalog --> Iceberg + Delta + Hudi + Hive through one catalog

Registry stats