What Is a Lakehouse Stack?
8 entriesA data lakehouse puts warehouse-style capabilities on open data in object storage. Its technologies fall into layers: object storage holds the bytes; file formats such as Parquet organise them; an open table format adds a metadata layer that provides tables, ACID transactions, schema evolution and time travel; a catalog tracks tables and governs access; and query engines read and write them. Because each layer is open, you can mix engines and vendors, which is the main difference from a closed warehouse. The idea was set out in the CIDR 2021 lakehouse paper.
Storage & File Formats
12 entriesObject stores hold the data, and columnar files make it efficient to scan.
Open Table Formats
18 entriesThe metadata layer that turns files into tables. Iceberg's specification lists four format versions: v1 analytic tables, v2 row-level deletes, v3 extended types and capabilities (nanosecond timestamps, variant, geometry and geography types, default column values, multi-argument transforms, row lineage, binary deletion vectors and table encryption keys), and v4, which restructures metadata with an adaptive metadata tree for efficient small commits and column updates, and combines data files with their deletion vectors. DuckLake takes a different approach by keeping metadata in a SQL database.
Catalogs & Governance
15 entriesCatalogs track tables and enforce access; the Iceberg REST catalog API gives engines a common interface.
Interoperability
3 entriesTools that let one set of data files be read through several table formats.
Query & Compute Engines
12 entriesEngines that read and write lakehouse tables.
Ingestion, Streaming & Transformation
10 entriesGetting data into the lakehouse and shaping it.
Architecture Patterns & Learning
3 entriesHow teams organise lakehouse data and learn the stack.
Related Pages
4 entriesNeighbouring pages on this site.
Frequently Asked Questions
What is a data lakehouse?
An architecture that puts warehouse-style capabilities, such as ACID transactions, schema management and fast SQL, on open data held in object storage, using an open table format and a catalog.
What is an open table format?
A metadata layer over data files that provides tables, transactions, schema evolution and time travel. Iceberg, Delta Lake, Hudi and Paimon are the main formats, and DuckLake keeps its metadata in a SQL database.
What does Iceberg v3 add?
According to the Iceberg specification, v3 adds new data types (nanosecond timestamps, variant, geometry and geography), default column values, multi-argument transforms, row lineage, binary deletion vectors and table encryption keys. Version 4 restructures metadata with an adaptive metadata tree.
Why do catalogs matter?
The catalog decides which tables exist, which engine may read or write them, and who has access. The Iceberg REST catalog API lets engines share one catalog, and projects such as Polaris, Unity Catalog and Nessie implement or complement it.
How do the lakehouse formats interoperate?
Tools such as Apache XTable and Delta UniForm let data written in one format be read through another, though feature support varies, so test your engines.
How current is this page?
Links were checked in October 2026. Technical statements about the Iceberg specification come from its own specification page (v3 and v4 sections), and DuckLake and Polaris facts from each project's own site; other descriptions summarize each project's pages. Open table formats, catalogs and engines are evolving quickly and interoperability features differ by version and vendor, so confirm support on each project's documentation. This page covers the underlying technologies; the Data Lakehouse Platforms page covers commercial platforms. Listing is not an endorsement. Sites that block automated checks, such as Dremio and Presto, are kept and may need a manual click.
What has changed in lakehouse technologies in 2026?
DuckDB 2.0.0 was released on 21 October 2026, Apache Iceberg 1.12.0 on 30 September, Delta Lake v4.4.1 on 5 October, Apache Arrow 26.0.0 on 9 October and Polars py-2.0.0 on 6 October. Apache Polaris graduated to an Apache top-level project on 19 February 2026. The MinIO repository was archived on 25 April 2026 and says it is no longer maintained, with the community edition distributed as source code only, so teams using it for lakehouse storage should plan a migration or move to the vendor's AIStor editions or another S3-compatible store. SAP agreed on 4 May 2026 to acquire Dremio, with closing expected in the third quarter of 2026.