Guide, read 10 Nov 2026
What is a data lakehouse
All guidesOpen the Compute Bill
The definition the term started with
The word lakehouse entered the research literature in "Lakehouse: A New Generation of Open Platforms that Unify Data Warehousing and Advanced Analytics" by Michael Armbrust, Ali Ghodsi, Reynold Xin and Matei Zaharia. It appeared at the 11th Annual Conference on Innovative Data Systems Research (CIDR), held online January 11 to 15, 2021. The authors list Databricks, UC Berkeley and Stanford as their affiliations.
The paper defines the pattern by three traits: open direct-access data formats such as Apache Parquet and ORC, first-class support for machine learning and data science workloads, and performance the authors describe as competitive with cloud data warehouses. That last trait is the authors' own claim, and the paper treats SQL performance on open files as its largest open technical question.
Databricks's own glossary page gives a looser, vendor-written version: a lakehouse combines the flexibility, cost-efficiency and scale of data lakes with the data management and ACID transactions of data warehouses. Both definitions agree on the mechanism. The data stays in files you can read without the vendor's engine, and a metadata layer decides which files make up a table.
Where the idea came from
The paper describes three generations. In the first, data was loaded into a warehouse with schema-on-write, and compute and storage were coupled in on-premises appliances. In the second, raw data was offloaded into data lakes holding open files, starting with Hadoop and HDFS, with a small subset later copied into a warehouse for business intelligence.
From 2015, the authors write, cloud object stores such as S3, ADLS and GCS replaced HDFS. They call the resulting pair of a cloud lake plus a downstream warehouse the two-tier architecture and say it was dominant in their experience. The paper cites a survey by Dimensional Research and Fivetran stating that 86% of analysts use out-of-date data. That survey figure is the paper's citation, not something this site measured.
The lakehouse is the proposal to collapse the two tiers: keep one copy on the object store, and give it the management features that made warehouses useful, which the paper lists as ACID transactions, versioning and indexing.
Warehouse, lake and lakehouse side by side
| Trait | Data warehouse (paper's first generation) | Data lake (second generation) | Lakehouse (paper's proposal) |
|---|---|---|---|
| Write model | Schema-on-write | Schema-on-read | Files plus a transactional metadata layer |
| Storage | Compute and storage coupled in on-premises appliances | HDFS, then cloud object stores | Cloud object store, open formats such as Parquet |
| Transactions and versioning | Built in | Left to the user, per the paper | Provided by the table layer (Delta Lake, Iceberg, Hudi) |
| Machine learning access | Export files or read over ODBC or JDBC | Direct file reads | Direct file reads with table semantics |
| Copies of the data | One | Lake plus a warehouse copy | One, per the proposal |
What a table format adds
The paper's proposed implementation is a transactional metadata layer over an object store that defines which objects belong to a table version. It names Delta Lake, Apache Iceberg and Apache Hudi as systems in this area. It says Iceberg started at Netflix and Hudi at Uber, and it states that Delta Lake was then used in about half of Databricks's workload by thousands of customers.
The paper also lists limits that still shape operations today: object stores have high latency, which caps the transaction rate, and Delta Lake, Iceberg and Hudi support transactions on one table at a time. Separately, data files on an object store are written once, so every write and compaction leaves older files behind until a cleanup job removes them. That is why the maintenance pages exist on this site.
Performance evidence in the paper
The paper reports one experiment: the Databricks Delta Engine, a SQL engine over Parquet, against popular cloud data warehouses on AWS, Azure and Google Cloud, using TPC-DS at scale factor 30K, with time and cost shown in its Figure 3. The authors ran it and they work at Databricks, so it is a vendor-run result. It is a SQL query benchmark, not a Spark batch ETL test, and this site does not restate its numbers.
Read it as the origin of the claim that open-format files can match warehouse speed, not as a comparison of today's platforms. Later vendor results are listed on each platform page with who ran them.
What the term leaves open
Vendors now use the word for different products. Databricks applies it to its platform, Snowflake and BigQuery offer Iceberg tables on cloud storage, and AWS assembles Iceberg or Delta on S3 with EMR and Glue. The formats pages separate the table format from the platform that runs it.
Questions that the definition does not answer, and that published facts can: which engines can write the table, who runs compaction, where the catalog lives, and what moving the data out would involve. The guides on choosing a table format and leaving a managed platform take those in turn.
Sources
- Armbrust, Ghodsi, Xin, Zaharia, "Lakehouse", CIDR 2021, read 10 Nov 2026
- Databricks, What is a data lakehouse, read 10 Nov 2026
- Apache Iceberg home page, read 10 Nov 2026
- Delta Lake home page, read 10 Nov 2026
Last reviewed 10 Nov 2026.