Common usage calls a data warehouse an analytical database where data only enter once they have been put into the format of a model defined in advance. It calls a data lake a storage space where data are deposited in their original format, with structure only applied on reading. Asking “data warehouse or data lake” therefore amounts to choosing the moment when the organisation pays for the structuring effort: before loading, once for all readers, or on reading, use by use.
The warehouse prevails when the questions are known and several departments need to read the same figures. The lake prevails when the data come in varied formats or their use has not yet been decided. From this difference in structure follow three others, concerning the uses served, governance and the shape of the cost. It also leaves room for a third way, the lakehouse, and for the coexistence of the two systems.
Structure is fixed on writing or on reading
This vocabulary is found in vendors’ documentation. Microsoft’s (Azure Architecture Center, page “What is a data lake?”, consulted on 12 November 2026) contrasts the lake’s schema-on-read with the warehouse’s schema-on-write. Amazon Redshift’s documentation, read the same day, presents this service as a managed data warehouse.
With schema-on-write, a load that does not comply with the model is rejected or corrected before reaching readers, and all of them query the same table. The trade-off is paid with each new item: one more source or field requires modelling work before it is available, and what the model did not anticipate does not get in.
With schema-on-read, deposit waits for nothing. Each processing job gives the files the structure it needs, which shifts control onto each reader.
Take the operator of a ski resort’s lifts. Its ski pass sales produce regular rows: pass type, point of sale, date, price. Put into the warehouse model, they give revenue per day and per pass type, which operations and finance read identically. Gate passages, wind readings from the lifts and maintenance reports written in free text have neither stable columns nor a settled question: the lake keeps them as they are, until the day a team wants to relate lift stoppages to weather and visitor numbers.
Uses that only partly overlap
The comparison of uses is best started with the readers. The warehouse is aimed at readers who query in SQL or from a reporting tool, without writing processing jobs: management controllers, business line managers, analysts. It owes them stable tables and indicators calculated only once, the subject of our business intelligence offering.
The lake is aimed at readers capable of programming the preparation they need: data science teams who explore a detailed history or train a model, engineers who prepare the warehouse tables.
The two mistakes are symmetrical. Opening a raw lake to readers who expect a ready-made figure sends them back to their extracts. Imposing the warehouse model on data that nobody yet knows what to do with consumes modelling effort with no one asking for it.
Governance and cost change place
In a warehouse, the model does part of the governance work: a table has a name, typed columns, access rights. It does not do the rest, since it says neither who decides the definition of an indicator nor who authorises access.
In a lake, nothing is described until someone has declared it. Each deposited dataset calls for an entry in a data catalogue, an owner and a retention rule, failing which the lake turns into the data swamp described in our data lake entry. Naming these owners falls, for the warehouse as for the lake, under data governance.
Cost follows the same shift. The warehouse makes you pay first: modelling and loading pipelines precede the first dashboard. The lake makes you pay later and in a diffuse way: deposit asks little, but preparation is redone by each reader, volumes grow as long as nothing is purged, and each read consumes compute. We do not consider either one less costly in principle. The comparison is made on actual uses, attaching spending to each, according to FinOps practices.
Data warehouse or data lake depending on the starting situation
| Starting situation | What prevails | Reason |
|---|---|---|
| Settled management indicators, read by several departments | Warehouse | The calculation rule is fixed once, before loading |
| Data in rows and columns, moderate volumes, a small team | Warehouse | A single system to operate, with SQL skills |
| Texts, images, application logs or sensor readings to retain | Lake | These formats are deposited without a prior model |
| A detailed history whose use remains to be defined | Lake, under a catalogue | The structure will be chosen along with the question |
| Models to train on non-aggregated data | Lake, or lakehouse | The raw detail remains accessible to processing |
| Both populations of readers on the same data | Both systems, or a lakehouse | The next section compares these two answers |
An organisation may recognise itself in several rows at once. The table then designates, use by use, the system that carries each need, without naming an overall winner. Choosing the model, then building it in slices of use, is the subject of our data platform offering.
The lakehouse and coexistence, two answers to “both”
When both needs exist, the first answer is coexistence: the lake receives the raw data, and processing derives the warehouse tables from them. Its trade-off: two datasets to keep in step and, for each figure, a reference system to designate.
The second is the lakehouse: managed tables, with transactions and controlled schemas, placed on the lake’s storage by means of open table formats such as Apache Iceberg or Delta Lake. Our entry details its components and limits. The labels are also converging: the Snowflake and BigQuery documentation, consulted on 12 November 2026, each describe tables in Apache Iceberg format whose files can reside in storage administered by the customer.
The lakehouse does not remove the initial choice: it moves it inside a single foundation. When it is organised in layers, its raw layer can remain structured on read while its exposed layers are structured on write, and the decision is made again layer by layer. Among the limits the entry notes, two weigh on the choice: the engineering skills to bring together, and an oversized foundation when the need is limited to management reporting.
The preliminary work can start without any tool: draw up the list of uses expected within the year and note, for each one, the data format, the readers and the question asked. Compared with the table above, this list says which rows concern the organisation, and in what order to serve them.