What is a schema?
A schema is the blueprint for a dataset: which tables or record types exist, which fields each contains, what type each field holds, which are required, and how records in one table refer to another. A relational database enforces the schema and rejects a row with text in a date column. In a JSON document store or a data lake, the schema may be applied only when data is read, which is flexible but error-prone. A canonical schema is the company-wide version used to translate between systems. Schemas change as the business adds and retires fields, and managing that change is one of the harder parts of running a data platform.
Why a schema matters
A schema is the shared vocabulary between those who write data and those who read it. A credit union’s member table with a typed date of birth, a required member identifier, and a foreign key to the branch table can be joined, validated, and governed, while the same facts in free-text notes cannot. For AI, schemas matter twice: structured outputs from a language model are checked against a schema before being written to a system of record, and text-to-SQL assistants depend on readable table and column names to produce correct queries.
What to watch for
- Name tables and columns for humans, since every analyst and AI assistant will read them.
- Version schema changes and announce them, ideally through a data contract.
- Decide where schema enforcement happens for semi-structured sources, since applying it late pushes errors onto every consumer.