Notes from inside other people’s estates

An invoice tells you what you spent, never what you bought. The gap between the two is where most of the savings are.

Almost every dollar wasted in a warehouse falls into one of four patterns. Each has a different fix and a different risk of breaking something.

Two warehouses, two billing models, two definitions of a unit of work. Comparing them honestly is harder than any vendor's calculator suggests.

Unused reports are not a reporting problem. They are the visible end of a chain of models, tests and jobs that still run every night.

The same breakage every night should be one thing with a count, not thirty notifications. Grouping by the shape of an error changes what a team can see.

A patch that compiles is not a fix. The evidence that a change is correct is more valuable than the change, and harder to produce.

Every mature estate contains assertions that were weakened under deadline and never restored. Finding them is straightforward; the hard part is the conversation.

Upstream changes rarely break anything immediately. They break something three weeks later, in a report, in front of somebody senior.

"Nothing reads this" is a claim about every consumer you have, including the ones you forgot. Here is what it takes to make it safely.

Table-level lineage answers "what breaks". Column-level answers "what changes", which is a different and usually more urgent question.

Data catalogues fail for the same reason wikis do. The fix is to stop treating the catalogue as a destination.

A surprising amount of tooling assumes the project sits at the root of the repository. Real repositories rarely oblige.

An agent that can answer questions about your data does not need to see your data. Drawing that line early decides what you can safely automate.

The constraint sounds like caution. It is actually what makes the rest of the automation acceptable.

Treated as a formality it becomes a rubber stamp. Treated as a measurement it tells you exactly how much autonomy to grant.

The hardest behaviour to build into an automated analyst is the refusal to produce a number it cannot support.

No error, no failure, no alert. Just a figure that is different from last week and somebody who needs to know why by lunchtime.

Data teams are asked to justify themselves in numbers they do not have. A few are genuinely measurable; most of what gets reported is not.

Small data teams inherit estates sized for the organisation, not for the team. The failure modes are identical to a large estate; the slack is not.

Most of what a governance initiative is meant to produce can be derived rather than documented — and derived facts do not go stale.