Most analysis problems turn out to be data problems wearing a disguise. Before a dashboard can be trusted, the pipeline behind it has to be boring: predictable schedules, explicit schemas, and failures that announce themselves rather than quietly producing a smaller number.
Model before you move
We agree the model — entities, keys, grain — before writing the first pipeline. Reshaping a warehouse after it has consumers is an order of magnitude more expensive than getting the grain right at the start.
Make quality a test, not a habit
Row counts, null rates, referential integrity and freshness are assertions that run on every load. A pipeline that fails loudly is worth several that succeed quietly with the wrong numbers.
