Dataset versus analysis
A dataset is prepared, reusable data — joins, field types, dataset calculations, row-level security. An analysis is where you build visuals on top of it.
Work done in the dataset is shared by every analysis using it. Work done in an analysis belongs to that analysis alone.
So: anything another report would also want belongs in the dataset.
Joins
The dataset editor joins tables visually — inner, left, right and full outer, on one or more field pairs.
The join diagram shows the configuration, and you can preview the result before saving. As everywhere, watch for a one-to-many join multiplying rows and inflating measures.
Joins across different data sources are supported, which is genuinely useful — a Redshift table joined to an uploaded CSV of targets.
Dataset filters and calculations
Filters applied at dataset level apply to every analysis built on it — the right place for excluding test accounts or restricting to relevant years.
Calculated fields defined here are equally shared. Define a Margin or Net Revenue field once in the dataset and every analysis uses the same definition, which is how two reports avoid disagreeing.
Field types and roles
Set data types, and change a field's role between dimension and measure. Numeric identifiers arrive as measures and get summed unless you change them, exactly as in every other BI tool.
Geographic roles — country, state, city, postcode, latitude/longitude — must be set for maps to work.
Rename fields to what a business user would call them. The dataset is what report authors see.
Excluding fields
Unselect fields you do not need. In SPICE they consume capacity; in direct query they widen every query. It is one click here and saves work everywhere downstream.