Glossary
Data lake
A data lake is a central repository that holds large volumes of data in its original format, structured or not, with no schema imposed when data is written. It usually runs on object storage and keeps storage separate from the engines that query or train on it.
Lake and warehouse compared
| Data lake | Data warehouse | |
|---|---|---|
| Data held | Raw and prepared data of any type | Curated, structured records |
| Structure applied | When data is read | Before data is loaded |
| Engines | Many engines read the same files | Tied to the warehouse's own engine |
| Typical use | Exploration, AI training and retrieval, long retention | Reporting and repeatable analytics |
Many organizations run both and feed the warehouse from the lake. The lake also keeps data the warehouse never accepted, such as images, audio, logs and documents, which is what makes it the usual starting point for AI work. Because storage and compute are separate in a lake, a new query or training engine attaches to existing data, and a retired one leaves the data untouched.
How does a lake turn into a swamp?
The usual failure is a lake holding several petabytes in which no one can find a usable dataset. Storage is not the cause. Data that lands without an owner, a source or a description cannot be found later by anyone who was not there when it arrived. The freedom to write anything is also the freedom to write it unlabeled.
The countermeasure is recording metadata at ingestion: owner, source, a few schema hints and a retention date. Reconstructing that years later depends on memory. A lake with no deletion rule also grows without limit, and joins the cost pattern described under storage sprawl. A data fabric depends on the same catalog discipline.
Why object storage is the usual base
A lake needs low cost per terabyte, growth into petabytes and simultaneous access from many engines. Object storage supplies all three, and S3-compatible storage gives nearly every analytics and AI tool a common way to read it. That shared base is why a lake often becomes the store stage of an AI data pipeline.