Analytics & BI Tools · Chapter 3 of 18
Tableau, in practice
The strongest visual analysis tool, and the one whose cost model surprises people.
Tableau's differentiator has always been exploratory visual analysis — the experience of dragging fields and seeing the shape of the data immediately. If your users are genuine analysts who think visually, it is still the best in class at that.
What it genuinely does well
- Exploratory analysis. The interaction model is designed for asking the next question, not for building a fixed report. Nothing else feels quite the same.
- Visualisation quality and control. More chart types, finer control, better defaults for perceptual accuracy.
- Analyst community and craft. A large body of practice and a genuine culture around visual design.
- Cross-platform — less tied to one vendor's ecosystem than Power BI.
Where it hurts
Cost. Per-creator licensing is materially more expensive than Power BI, and the gap widens with scale. This is the single most common reason organisations move away, and it is rarely a capability judgement.
Governance and semantic modelling are weaker. Extracts and workbook-level calculations make it easy for every workbook to define metrics differently — the exact failure from chapter 1.
Extracts drift. The performance answer is usually an extract, which is a copy with a refresh schedule and therefore a staleness and storage problem.
Enterprise deployment is heavier than the sales conversation implies, particularly around permissions at scale.
When Tableau is the right answer
When your users are analysts doing genuine exploration rather than consuming fixed dashboards, when visual quality matters to the audience, and when you have the budget and a governance plan for the semantic layer.
When it is not: a large population of casual consumers reading fixed reports. That is expensive in Tableau and cheap almost everywhere else.
The order of operations, which explains every surprising number
Almost every ‘why is this number wrong?’ question in this tool is answered by the order in which filters and calculations are applied. The order is fixed, documented, and almost never taught — which is why so many analysts learn it by trial and error.
| Stage | What happens | What it means in practice |
|---|---|---|
| 1. Extract and data-source filters | Applied before anything else | The cheapest place to filter, and invisible to everything downstream |
| 2. Context filters | Create a temporary subset that later steps operate on | This is what makes a top-N filter work with another filter. Without context, the top N is computed against the whole set |
| 3. Fixed-level expressions | Computed at a stated grain, ignoring the visual's dimensions | Where percentage-of-total and cohort calculations live |
| 4. Dimension filters | The ordinary filters most people use | Applied after fixed expressions, which is why a fixed calculation does not change when you filter |
| 5. Include/exclude expressions | Computed relative to the visual's grain | Useful for average-of-an-aggregate, and the source of much confusion |
| 6. Measure filters | Filter on aggregated values | This is why filtering on a measure behaves unlike filtering on a dimension |
| 7. Table calculations | Computed on the result set already returned | They cannot see filtered-out data, which is why a running total changes when you filter |
The two rules that resolve most confusion: a fixed-level expression is computed before ordinary dimension filters, so filtering does not change it — which is either exactly what you wanted or exactly your bug. And a table calculation is computed after everything, on the rows that survived — so it is a presentation-layer operation and can never recover data a filter removed.
The practical diagnostic: if a number changes when you filter and you did not expect it to, you are probably using a table calculation where you wanted a fixed expression. If a number stubbornly refuses to change when you filter, you are using a fixed expression where you wanted a table calculation, or you need a context filter.
Check yourself