Lakehouse tables
Open lakehouse table formats are first-class citizens, not a secondary export path. Delta and Iceberg ship together, co-equal, and the table format is a deployment policy a cell chooses from configuration, not a rewrite and not a house format a customer is locked into.
In one paragraph
We publish and share in open formats, bring your own engine. Delta and Iceberg are co-equal and the table format is a deployment policy; both read zero-copy over object storage; and external sharing runs over both the Delta Sharing protocol and the Iceberg REST catalog, under one governance model.
Format neutrality
Format neutrality is the differentiator here, and it is concrete:
- Both formats ship as core dependencies. The boot path prefers neither.
- A single policy setting chooses the default, and a per-table choice overrides it. An Iceberg-primary sovereign cell boots from configuration alone, no code change.
- Both read zero-copy over object storage: the cell's own object store, or external S3, GCS, or Azure, so a reader is not paying a copy to consume a table.
A cell writes and merges tables in both formats, with a format-agnostic core underneath (commits, deletion vectors, predicates, table clone) carrying the format-specific layers above it.
*Figure C6: one lakehouse core, two co-equal formats. A single deployment policy setting chooses the default format (a per-table choice still overrides it); Apache Iceberg and Delta Lake sit co-equal on one format-agnostic core, joined by a bidirectional metadata bridge so either is queryable through the other; and both read zero-copy over object storage, the data stays where it lives.
Table equivalent
| Element | Role | Notes |
|---|---|---|
| Deployment policy setting | Chooses the default table format | Per-table override allowed; boots from configuration alone (no code change) |
| Format-agnostic core | Carries commits, deletion vectors, predicates, table clone | The neutral layer beneath both formats |
| Apache Iceberg | Co-equal open format | Read and write |
| Delta Lake | Co-equal open format | Read and write |
| Bidirectional metadata bridge | Delta ↔ Iceberg metadata conversion | Either format queryable through the analytical path |
| Object storage | Where the bytes live | Zero-copy read: cell store, S3, GCS, or Azure |
What each format brings
| Format | Notable surface |
|---|---|
| Apache Iceberg | Schema fields and partition transforms, sort orders, snapshot management (create, expire, rollback), schema and partition evolution, manifest and data-file management across Parquet/ORC/Avro, table maintenance, and a REST metadata catalog. |
| Delta Lake | A JSON transaction log and version management, branch management for concurrent workflows, checkpoints, row lineage, variant type for semi-structured data, native geospatial types, and the Delta Sharing protocol. |
| Interoperability | Bidirectional Delta ↔ Iceberg metadata conversion and analytical-engine bridges, so either format is queryable through the analytical path. |
Honesty about scope
The claim to make here is format neutrality and zero-copy read parity between the two open formats, plus open sharing. Two claims sit deliberately outside that ceiling:
- Format neutrality, precisely. The claim is neutrality between the two open formats, Iceberg and Delta at parity, which is a different and more useful property than matching any single vendor platform feature for feature.
- Not a performance claim. The acceleration layer that makes a lakehouse fast, file pruning, row-group statistics, Bloom filters, Z-ordering, projection indexes, and related data-skipping, is built but is not in the production read path, so no performance or benchmark claim attaches to the open formats.
Nothing in the platform is graded, certified, or deployed against a reference customer; these pages describe mechanisms, not proven results.
Across the deployment matrix
- Single-cell and multi-cell, tables live in the cell's stores and read zero-copy locally or across cells.
- Sovereign / air-gapped, an Iceberg- or Delta-primary sovereign cell is a configuration choice; no external service is required to read a table.
- Cloud-enabled: external object stores are available as an option.