query-design

Query Design — Separation of Data and Presentation

DAX computes and fetches data. TypeScript shapes it for display. VegaVisual and DataGrid render it.

Aggregate in DAX to the visual's grain — never fetch lower-grain rows to roll up client-side. When a visual layout changes, only the TypeScript or spec layer should change — not the DAX query.

Responsibility Matrix

Rules

Must

• Aggregate in DAX to the visual's grain — never fetch lower-grain rows just to roll them up to that grain in TypeScript
• One EVALUATE per .dax file; use ROLLUPADDISSUBTOTAL when DataGrid body and grand-total grains share one result table
• ORDER BY in DAX for stable, diffable results — not presentation sort
• Same filters/measures across related split-grain queries to prevent drift

Prefer

• SUMMARIZECOLUMNS for grouped aggregation — it also drops BLANK-measure rows, keeping payloads small
• DAX's natural column names ('Table'[Column], [Measure]) mapped via columnMetadata.displayName
• Raw typed values from DAX — format via columnMetadata.format or Vega-Lite, never FORMAT()
• Model-defined format strings (from INFO.VIEW.MEASURES()) over invented ones
• Multiple lightweight queries for independently shaped datasets; one flagged rollup query for a DataGrid body and grand total
• User-facing sort in TypeScript / Vega-Lite / DataGrid — never re-query for sort

Avoid

• SELECTCOLUMNS solely for renaming — use columnMetadata.displayName instead
• UNION to mix body and total grains — use ROLLUPADDISSUBTOTAL and partition its flagged rows
• FORMAT() in DAX — converts to text, breaks sorting and charting
• Converting BLANK to 0 / "" / "N/A" in DAX — causes result-set explosion
• CONCATENATEX, UNICHAR, emoji prefixes — decorative text belongs in cellRenderer or Vega-Lite
• Fetching all members of high-cardinality dimensions just to fill gaps

Decision Flowchart

Need to add something to the query result?
  |-- Calculation / aggregation / filter?
  |     -> DAX (measures, CALCULATE, SUMMARIZECOLUMNS)
  |-- Interactive filter the user controls?
  |     -> Low-cardinality: widen grain, filter in TypeScript or Vega-Lite transform
  |     -> High-cardinality: push filter to DAX, re-query
  |-- Adding a DataGrid grand total?
  |     -> DAX rollup query: split body and total rows with toRollupDataTables
  |        and always pass the returned grandTotalTable via grandTotals.data
  |     -> DataGrid-computed sum/count: pass grandTotals without data and set defaultAggregation
  |-- Merging datasets or adding other synthetic rows?
  |     -> Charts: pass multiple DataTables to VegaVisual, layer in spec
  |     -> Grids: append rows in TypeScript, style via cellRenderer
  |-- Renaming a column for display?
  |     -> columnMetadata in the factory file (displayName)
  |-- Formatting, labeling, or encoding?
  |     -> Vega-Lite spec or DataGrid cellRenderer
  |-- Decorating values (icons, status badges, null placeholders)?
  |     -> DataGrid cellRenderer or Vega-Lite condition encoding
  |-- Not sure?
        -> Does it change what the data *means* (filter, measure, grain)? -> DAX
           Does it change only how data is *rendered* (labels, icons, layout)? -> TypeScript / Vega-Lite spec / DataGrid cellRenderer
           Still unclear? -> Read the relevant reference above

Interactivity

Reports coordinate multiple visuals: a selection in one changes what the others show. Two distinct behaviors, with different data work behind them:

• Cross-filtering — a selection in one visual constrains the data shown in another, removing or narrowing the non-matching rows from the target's view. The target shows less. Applying that constraint is a cost/cardinality tradeoff — widen the grain and filter client-side, or push the filter into DAX and re-query. See Filter strategy.
• Cross-highlighting — a selection in one visual emphasizes the matching subset within another while the full context stays visible. The target keeps its baseline (dimmed) and draws the selected subset bright on top. The subset is a separate aggregation aligned to the baseline's grouping, measures, and row set — not a client-side filter of the baseline. See Highlight queries.

Both consume the predicate-based selection events the visual components emit (onInteraction). The components render only the DataTables they are handed; this skill produces those tables. For how a spec binds and layers multiple datasets, see the visuals skill's multi-data input reference.

Reference Materials

Read these when working on a specific topic:

• Anti-patterns and corrections — Open when reviewing a query that uses UNION for totals, FORMAT(), SELECTCOLUMNS for renaming, CONCATENATEX/emoji decoration, BLANK-to-0 conversion, or GENERATE/CROSSJOIN for gap-filling.
• Multi-grain patterns — Open when a single visualization needs data at multiple grains (e.g., bars + reference line, region detail + total row, monthly trend + YTD), to choose between one flagged rollup result and separate queries while keeping related factories, hooks, and DataTables aligned.
• Filter strategy — Open when adding a user-controlled filter or implementing cross-filtering, and deciding whether to widen the grain (filter client-side) or push the filter into DAX (re-query on each change).
• Highlight queries — Open when writing the "selected subset" overlay query for a cross-highlight visual: an aligned CALCULATETABLE / TREATAS query whose rows match the baseline.
• Format strings — Open when picking a columnMetadata.format value, when a measure has a dynamic format string, or when formatting needs to flow into a Vega-Lite axis.

Integration with Sibling Skills

• schema-discovery — Schema exploration; discover tables, columns, and relationships before writing queries.
• dax-authoring — DAX syntax, query patterns, and testing workflow. Apply this skill's principles when deciding what DAX should compute.
• visuals — Vega-Lite specs and DataGrid configuration. Push formatting and labels into specs, not DAX.