In Tableau, in what order are extract, data source, context, dimension and measure filters applied?

## The ladder Tableau publishes an order of operations, and it is the single most useful thing to memorise about the tool. The filter-relevant stages, top to bottom: 1. **Extract filters** — evaluated when the `.hyper` extract is created or refreshed. 2. **Data source filters** — attached to the connection, applied to every worksheet that uses it. 3. **Context filters** — the ones you right-click and *Add to Context*. 4. **FIXED level-of-detail expressions**. 5. **Dimension filters** — including Top N, conditional filters and computed sets. 6. **INCLUDE / EXCLUDE level-of-detail expressions**. 7. **Measure (aggregate) filters**. 8. **Table calculations**, then **filters on table calculations**. 9. **Forecasts, trend lines and reference lines**, drawn from what is left. Each stage sees only what the stage above it passed down. That is the whole model, and nearly every "why is this number wrong" question in Tableau is answered by locating two things on this ladder and noticing which one runs first. ## Where the stages land in the query For a live connection, the top stages become SQL. Data source, context and dimension filters become `WHERE` predicates; measure filters become `HAVING`, because they can only be evaluated once the `GROUP BY` has run: ```sql SELECT region, SUM(sales) FROM orders WHERE order_date >= DATE '2024-01-01' -- dimension filter GROUP BY region HAVING SUM(sales) > 100000 -- measure (aggregate) filter ``` Table calculations are different in kind: they are computed by Tableau on the aggregated result set that comes back, not by the database. That is why they sit below everything the database did. ## Consequences you should be able to name **Extract filters are destructive.** Rows they exclude are not in the extract, so no worksheet can ever recover them, and a viewer with permission to see them still cannot. That is a feature when you are trimming an extract for size, and a trap when someone later needs last year's data. **Context changes what Top N sees.** A Top 10 filter and a `Region` filter are both dimension filters, so they run at the same stage and neither can see the other's result. Adding `Region` to context moves it above the Top N, so the top ten are computed within the region. **A FIXED expression ignores dimension filters.** It computes at stage 4, before dimension filters exist. If you need it to respect a filter, that filter has to be in context — or be a data source or extract filter. **A measure filter removes marks, not rows.** Filtering to `SUM(Sales) > 100000` at region grain removes regions. Change the view to city grain and the same filter now removes cities, and the totals move, because the aggregation the filter tests is computed at the view's level of detail. **A table-calc filter is a late filter.** If you filter on a running total or an `INDEX()` calculation, the calculation was already computed over the full set; the filter only hides marks afterwards. That is exactly how you build "rank within the filtered view" behaviour, and exactly why a percent-of-total does not re-scale when you filter that way. ## Context filters have a cost A context filter is not free. Tableau materialises the contexted subset, and whenever the context filter's value changes, that subset is rebuilt and everything downstream is recomputed. Putting a frequently-changed filter card into context — a date range a viewer drags constantly — makes every interaction pay for a rebuild. Context is for filters that are stable within a session and that something downstream genuinely needs to run *after*: Top N, computed sets, FIXED expressions. ## Choosing a stage on purpose Think of the ladder as a set of scopes. Something that must never be visible to anyone belongs in an extract or data source filter. Something that establishes the population a calculation reasons about belongs in context. Something the viewer changes belongs in an ordinary dimension filter card. Something that tests an aggregate is a measure filter by definition. And something that must not disturb an already-computed calculation belongs on a table calculation. ## What interviewers listen for They are not usually testing rote recall of all nine stages. They want you to say the order at least roughly, then apply it to a symptom: given "my top ten shows six rows" or "my FIXED calculation ignores my filter", the answer is the same sentence — those two things are on different rungs, and here is which one runs first.