What actually changed in this field in three years

We started this in January 2023. Enough has happened since to be worth an audit: which lessons aged, and which did not.

What did not age at all. Grain. Joins. Idempotency. Missingness. The difference between a correlation and a cause. Every lesson in the foundations, quality and statistics modules stands essentially as written, and I would expect the same in another decade. These are properties of data and of reasoning, not of tools.

What needed touching up. The tooling article. Some specific recommendations moved — defaults changed, a couple of tools consolidated, single-machine engines got substantially more capable and moved more workloads out of "you need a cluster" territory. The shape of the advice held: pick the least machinery that expresses your dependencies.

What we rewrote. The section on when to reach for a model. The economics changed enough that some problems previously worth a bespoke model are now better served by a general model behind an API, and some previously impossible things became routine. What did not change is the framing question — is there an action attached, and what is it worth — which turns out to be the durable part.

What we were wrong about. We underrated how much of the audience would be people using data tooling occasionally rather than professionally. The lessons assume you will write SQL. A growing share of readers mostly read SQL that something else wrote, and the skill that matters for them is verification rather than authoring. We have not solved this and it is the biggest gap in the course.

The general pattern, and it is not a surprise: the closer a lesson is to a tool, the faster it dated. The closer it is to a question — what does one row mean, what decision does this change, how much should I trust this — the more it held.

That is an argument for what to spend your own learning time on.