The material is attributed by practice area rather than by person. Nine disciplines, each owning the modules and posts closest to its day job — so every page tells you which kind of practitioner it came from.
Ingestion, orchestration, and the unglamorous parts that keep everything else standing.
Batch and streaming pipelines, idempotency, backfills, schema drift and data contracts. The recurring theme: a pipeline that cannot be rerun safely is a script with ambitions.
If your pipeline has no idea what to do when yesterday shows up twice, it is not a pipeline.
Turning warehouses full of raw tables into models an analyst can trust without asking anyone.
Dimensional modelling, grain, slowly changing dimensions, metric definitions, and the SQL that survives production. Most "we need a data scientist" problems are "we need one correct table" problems.
A metric without a definition is a rumour with a chart attached.
Designing tests that can succeed, and reading the results honestly.
Power and sample size, randomisation units, novelty effects, sample ratio mismatch, and what to do when you cannot randomise at all.
Most experiments do not fail. They are simply too small to have ever succeeded.
Uncertainty, sampling, and the difference between a result and a finding.
Describing a column without lying, confidence intervals, the bootstrap, multiple comparisons, confounding and Simpson's paradox.
The interval is the result. The point estimate is just where the interval happens to be centred.
Getting models out of notebooks and into places where they can do damage responsibly.
Framing, baselines, leakage, validation that matches deployment, feature engineering, interpretation — and the cases where the honest answer is not to build a model.
Your model is not in production until something wakes you up when it is wrong.
Tests, contracts, reconciliation, and finding out before the CFO does.
Testing data like code, the anatomy of a silent failure, and the routine for the first hour with an unfamiliar dataset. Also the deliberately broken dataset people seem to enjoy most.
Every dataset is dirty. The only question is whether you found out on your terms or theirs.
Storage formats, compute cost, and why your query costs forty dollars to run.
Columnar storage, partition pruning, file layout, query plans, orchestration and monitoring. Started as an internal wiki page titled "please stop using SELECT *".
Columnar storage is not a detail. It is most of the reason your query is fast or slow.
Charts that answer the question they were made for.
Choosing a form before choosing colours, colour as an encoding rather than decoration, and communicating a result to someone who will not read your notebook. Every tool on this site was built here.
If you need a legend, a key and a footnote to read the chart, the chart is doing the wrong job.
Turning a vague question from a stakeholder into an answerable one.
What the job actually is, scoping, grain, working setup, and knowing which decision an analysis is meant to change before anyone opens a query editor.
Nobody wants a dashboard. They want to stop worrying about something.