The dataset editor
The dataset editor is where preparation happens: join tables, exclude fields, rename, change data types, add calculated fields, and apply dataset-level filters.
Everything you do here is inherited by every analysis on that dataset. Removing unused fields is worth doing aggressively — it shrinks SPICE consumption and reduces clutter for authors.
Joins
The visual join editor supports inner, left, right and full outer joins across tables, and across different data sources in the same dataset.
As anywhere, a one-to-many join duplicates rows on the one side and inflates sums. Check row counts before and after. Where the warehouse can do the join, doing it in a view upstream is usually faster and easier to maintain.
Field configuration
Set data types explicitly — string, integer, decimal, date, boolean. Dates parsed as strings are a frequent cause of broken time filters.
Assign geographic roles (country, state, city, postcode, latitude, longitude) to enable map visuals. Without the role set, QuickSight will not offer a geospatial chart.
Set a default aggregation per measure so authors get a sensible starting point, and add field descriptions — they surface in the analysis and save a lot of questions.
Custom SQL
A dataset can be defined by custom SQL against the source rather than the visual editor. This gives full control — window functions, complex joins, source-specific syntax.
The trade-off is maintainability and that some QuickSight features work less smoothly on top of it. The usual recommendation is to put the logic in a view in the warehouse and point a simple dataset at that: version-controlled, reusable outside QuickSight, and easier to debug.