Storage and catalogs

Engine streams Arrow batches directly from registered lake data and resolves immutable worker inputs through its native catalog.

Reads

Local, ADLS Gen2, and S3 Parquet readers preserve requested column order and conservatively prune row groups. Delta supports contiguous JSON history and complete classic or multipart checkpoints, with coordinator-pinned versions. Empty Delta tables retain their declared schema. The read-only Iceberg v1/v2 path accepts a committed metadata JSON pointer and reads supported Parquet snapshots by field ID. Deterministic splits feed distributed scans; row-level filters preserve correctness.

Catalog

SQLite/WAL metadata provides transactions, migrations, stable IDs, optimistic revisions, structured Arrow schemas, lifecycle enforcement, credential references, and audit history. Workers use coordinator-resolved fragment locations.

Cloud readers have local object-store tests; live cloud deployment qualification remains pending. Delta partition reconstruction, column mapping, deletion vectors, and unsupported schema evolution fail explicitly. Iceberg active deletes and nested/default fields are unsupported. Platform and Engine catalogs remain separate stores connected through explicit revision-checked synchronization; external catalog adapters remain targets.