Ask three consultants for a data architecture and you will get a warehouse, a lake and a lakehouse, each defended with conviction. The technical distinctions are real but narrower than the debate suggests. The decisive differences are organisational.
What each is actually good at
- Warehouse. Structured, modelled, governed. Excellent when the questions are known and repeated — finance, regulatory reporting, sales performance. Schema discipline up front is the price and the point.
- Lake. Cheap storage for anything, in any shape. Excellent when you do not yet know what you will ask, or when the data is genuinely unstructured — images, logs, sensor streams. Without governance it becomes a place data goes to be forgotten.
- Lakehouse. Table formats over lake storage, giving transactions and schema management on cheap object storage. Increasingly the sensible default for a new build, at the cost of a younger toolchain and fewer people who have run it in anger.
The questions that settle it
Rather than comparing feature grids, answer these:
- Who writes the transformations — a central team, or analysts in each function? Distributed authorship needs far stronger contracts and testing.
- What is the reporting obligation? If figures are audited, lineage and reproducibility outrank flexibility.
- How much of the data is non-tabular today? If the honest answer is almost none, a lake is solving a problem you do not have yet.
- Who is on call when a pipeline fails at 6am before the board pack? Platform choice should follow the skills you can actually staff.
Cost behaves differently than expected
Storage is rarely the bill. Compute is, and compute is driven by how often things are rebuilt and how carelessly queries are written. A platform with elastic compute and no cost guardrails will produce an invoice that surprises the CFO in month four. Set budgets, alerts and query limits on day one, when nobody is under deadline pressure.
Governance is a product, not a policy document
A catalogue nobody updates is worse than none, because it is trusted and wrong. Ownership per dataset, freshness expectations published next to the data, and tests that fail loudly do more for trust than any written standard.
A pragmatic recommendation
For most mid-sized enterprises starting now: lakehouse-style storage with modelled, governed layers on top for the reporting that matters, and a modest number of well-owned datasets rather than a large number of orphaned ones. Buy the flexibility, but spend the discipline where the numbers are published.
