Data lake

A data lake is a central repository that stores data in its raw, native format without imposing a schema upfront. It accepts everything: structured tables exported from a database, semi-structured JSON, logs and event streams, and unstructured content such as documents, images, or audio. Structure is applied only at analysis time — an approach known as schema-on-read — which makes the lake a flexible foundation for analytics whose questions are not yet known when the data arrives.

How a data lake works

The storage layer is typically inexpensive, horizontally scalable object storage (S3-compatible) or a distributed file system. Data is organized into zones — raw, cleaned, and curated — and moves between them through ELT pipelines: load first, transform later. Query and processing engines such as Spark or Trino operate directly on the files, while a data catalog with metadata makes individual datasets discoverable and understandable. The contrast with a data warehouse is fundamental: the warehouse demands schema design before ingestion and accepts only processed data, while the lake ingests everything immediately and defers the cost of structuring to read time. The two models increasingly converge in the lakehouse architecture, which layers warehouse-style transactions and quality guarantees on top of lake storage.

Practical applications

Data lakes underpin most big data and machine learning initiatives:

• ML training — data science teams need raw historical data at full granularity, which warehouses rarely keep;
• log and telemetry analytics — event streams from applications, servers, and IoT devices land cheaply without losing detail;
• analytical archive — the complete history of source data available on demand for audits or reprocessing;
• source integration — one landing zone combining CRM records, transactional systems, and third-party feeds.

Governance decides whether the investment pays off: a maintained catalog, dataset ownership, access control, and quality rules. Without them, a lake degrades into a data swamp — terabytes of files nobody can interpret or trust, which is worse than having no lake at all.