Notes from the insights desk.

Short, practical articles for the people who pay for data systems and the people who run them. Every claim is sourced in the notes, and nothing here is sponsored.

Newspaper pages running through a printing press — Fig. 1Newsprint on the press. We try to write about data the way good papers report: checked, sourced, and readable.

Rows of white equipment cabinets in a data hall

Architecture 4 min read

What a modern data stack needs

Most teams need fewer tools than they think, and more of the unglamorous parts: tests, ownership, and documentation.

Read the article

Data quality 4 min read

Data quality checks that matter

Start with the checks that catch real incidents: freshness, volume, keys, and the rules your business already states.

Hand checking off items on a handwritten list

AI readiness 4 min read

Getting data ready for AI

Before a model comes history, labels, permissions, and a record of where every field came from.

Grid of wooden card catalog drawers with handwritten labels

B1.1Glossary

Terms we use, defined once.

If we use a word in a report, it means what it says here.

Person reading printed pages at a table — Fig. 2Definitions are for reading, so we keep them short.

Lineage: The recorded path a dataset takes from its sources, through each transformation, to the reports and models that use it.
Freshness: How recently a table was updated, compared with how current its users need it to be.
Grain: What one row in a table represents, such as one order, or one customer per day. Mixing grains is how totals double.
Staging: A first layer of models that renames, types, and cleans raw data without applying business rules.
Mart: A set of models shaped for one group's questions, such as finance or growth.
Change data capture: Copying only the rows that changed in a source, often by reading the database's transaction log.
Semantic layer: Shared definitions of metrics and dimensions that every report reads from, so revenue means the same thing everywhere.
Data test: A query that selects the rows breaking a rule. No rows means the test passes.
Orchestration: Running pipeline steps in the right order, with retries, schedules, and alerts.
Data contract: An agreement between the team that produces data and the teams that use it, covering structure, meaning, and quality.

Have a question you would like answered here?

Send it by email. If it is useful to others, we may write about it, without naming you or your company.

Email the desk