Columnar / Compare
Different layers.
Connected ideas.
These technologies answer different questions. Start with the responsibility you need to understand, then follow the concrete example.
Three questions to separate.
STORAGEHow are values laid out?
A file format organizes bytes, columns, pages and encodings. Parquet and Lance make this layer visible.
MEMBERSHIPWhich files belong?
A table format records versions and references. Iceberg, Delta and Lance use different metadata structures.
EXECUTIONHow is an answer computed?
An engine plans scans, joins and aggregates. DuckDB and DataFusion illustrate this work.
Responsibilities, side by side.
A technology can span layers. Lance combines a file container, table metadata and optional search structures; DuckDB includes both an execution engine and storage. This table describes responsibilities, not a ranking.
On a narrow screen, scroll the comparison sideways to see every layer.
What the examples actually demonstrate.
The Iceberg and Delta lessons start from the same weather records and both append and delete rows, but their deletes remove different observation IDs. It does not make their metadata interchangeable. The DuckDB/DataFusion queries read the same bounded weather file; they do not benchmark table-format performance.
Use each kind of evidence for its question.
File hashes and decoded ranges substantiate structure. Independent engine or reader results substantiate the demonstrated values. Recorded plans show selected operators. Conceptual playback explains a mechanism. None of these alone establishes a general speed or cost ranking.
Arrow has a dedicated buffer guide and IPC inspector, connected to the native result buffers in the engine guide. It is neither a table catalog nor a synonym for any on-disk format. The Lance guide compares exact search with a separate native IVF_FLAT index. Its original byte fixture remains unindexed; hosted LanceDB services are a separate product.
Follow one native table history in Iceberg and Delta →