Every second discovery call we run includes some version of this question. A finance director has read that they need a data lake. An ops manager wants "one place where all our numbers live". The honest answer is that the label matters far less than whether your team can keep the thing alive after we hand it over.
So before we talk products, let us define the two shapes properly and then get to the part nobody writes about: the running costs in people, not just infrastructure.
What each one actually is
A data warehouse stores structured, modelled data. Someone has decided what a "customer" is, what an "order" is, and how they join. Data arrives through pipelines that clean and shape it on the way in. Queries are fast and the numbers agree with each other because the modelling work happened up front.
A data lake stores raw files: CSV exports, JSON events, sensor logs, images, whatever you have. Nothing is modelled on arrival. The shape gets imposed later, at query time, by whoever needs the data. That flexibility is the whole point, and also the whole risk.
When a warehouse wins
If your questions are business questions, a warehouse wins almost every time. "What was margin by branch last quarter?" is a warehouse question. So is nearly everything a board pack contains.
- Your sources are mostly databases and SaaS tools: accounting, CRM, an ERP, a payments provider.
- Your consumers are analysts and dashboards, not data scientists.
- You care that two reports never disagree about revenue.
- You have one or two data people, not a platform team.
For a client like Lumina Retail, we landed point-of-sale and e-commerce data into a Postgres-based warehouse with dbt handling the transformations. Total build: nine weeks. One analyst maintains it in roughly four hours a week. That is the benchmark to beat.
When a lake earns its keep
A lake earns its keep when the data itself is messy, huge, or not tabular. TerraCore Resources feeds us vibration readings from mining equipment at about 40 million rows a day. Forcing that through warehouse modelling on arrival would be slow and expensive. It lands raw in object storage, and only the aggregates that people actually query get promoted into structured tables.
- You have high-volume event or sensor data measured in tens of gigabytes a day or more.
- You genuinely do machine learning work that needs raw history, not aggregates.
- Your sources include unstructured content: documents, images, call recordings.
- You have engineers who will own schemas, partitioning, and file hygiene.
The honest comparison
| Warehouse | Lake | |
|---|---|---|
| Data on arrival | Cleaned and modelled | Raw, as-is |
| Best consumers | Analysts, dashboards | Data scientists, ML pipelines |
| Time to first dashboard | 6 to 10 weeks | 12 weeks or more |
| Storage cost | Higher per GB | Cheap per GB |
| People cost | Low: 1 analyst part-time | High: engineering ownership |
| Failure mode | Stale pipelines | Unusable swamp |
The lakehouse middle ground
Modern table formats let you run warehouse-style tables directly on cheap object storage, which the industry calls a lakehouse. It is a real option, not just marketing. But it moves complexity around rather than removing it: you still need modelling discipline, and now you also need someone who understands file compaction and table maintenance. For teams under about eight engineers we rarely recommend starting here.
How we decide with clients
- 1List the ten questions the business most wants answered. If eight or more are tabular business questions, start with a warehouse.
- 2Add up daily data volume. Under 5 GB a day, a warehouse handles everything comfortably.
- 3Count the people who will own the platform. One part-time analyst means warehouse, full stop.
- 4Check for genuinely unstructured sources. If they exist, bolt a small raw-file store onto the warehouse rather than going lake-first.
- 5Revisit in a year. Warehouses grow into lakehouses gracefully. Swamps do not grow into anything.
The pattern that fails most often is the ambitious lake built for a future that never arrives. The pattern that succeeds is boring: a warehouse answering real questions within a quarter, extended only when a concrete need shows up.
If you are weighing this up for your own company, our data and analytics team runs a free scoping session where we work through exactly this checklist against your actual sources. Bring your ten questions.