Unit 3 — Preparation, Visualization and Storytelling with Tableau

Topics Covered

Syllabus topics: Introduction to Tableau; characteristics of Tableau; Tableau architecture and components — Tableau Public, Desktop, Reader, Online, Server; the Tableau interface — shelves, marks card, views; Tableau extensions; data connection and preparation — cleaning, pivoting, filtering; calculated fields and LOD expressions; basic visualizations — bar, line, tree, geo maps, scatter plots; storytelling with Tableau; creating a Tableau story. Case study — HR analytics.

3.1 What Tableau is, and its characteristics

Characteristics the syllabus wants named

3.2 Architecture and components

⚠️ Tableau Public makes everything public

Anything you save to Tableau Public is visible to the entire internet, and downloadable. It cannot save locally in the usual way — the workbook lives on their servers.

For labs with sample data that is fine, and experiments 8, 9 and 12 rely on it. Never put real student, employee or customer data in it. This is the single most important warning in the unit, and it is a genuine, repeated real-world data breach.

File types — a two-mark question

Hand in .twbx. A .twb on the examiner's machine opens with no data, and it is the commonest way to lose lab marks.

The architecture, for the diagram question

        Data sources (files, databases, cloud)
                        |
                  Data Connection
                (Live  or  Extract)
                        |
              +--- Tableau Desktop ---+          authoring
              |     (VizQL engine)    |
              +-----------+-----------+
                          | publish
        +-----------------+-----------------+
        |                 |                 |
  Tableau Server    Tableau Online     Tableau Public
   (self-hosted)     (SaaS)             (free, public)
        |                 |                 |
        +-----------------+-----------------+
                          |
              Browser / Mobile / Reader        consumption

Server's own components — Gateway, Application Server, VizQL Server, Data Server, Backgrounder, Repository — are worth naming if the question says "explain Tableau Server architecture". Backgrounder runs extract refreshes and Repository is the PostgreSQL database holding metadata; those two are the ones asked about.

3.3 The interface — shelves, marks card, views

⚠️ Blue vs green is the concept, not a colour scheme

This is the distinction that confuses everyone, and it is examined.

But it is not fixed by field type. A date can be either: discrete YEAR for one header per year, continuous for a real time axis. Converting between them changes the chart entirely, and "why did my line chart become bars?" is nearly always this.

Tableau extensions

Dashboard extensions are web applications embedded in a dashboard zone that can read and write the dashboard's data — write-back forms, custom filters, third-party visuals from the Extension Gallery. Analytics extensions connect Tableau to an external engine — Python via TabPy or R via Rserve — so a calculated field can call a model.

TabPy is the bridge to Machine Learning: train a model in scikit-learn, expose it through TabPy, and a Tableau calculated field can score rows live. Worth naming in the viva; rarely needed at this level.

3.4 Data connection and preparation

Cleaning, specifically

The syllabus names cleaning, pivoting and filtering as three things, so answer them as three. Cleaning in Tableau happens on the Data Source tab, and these are the tools:

⚠️ An alias is not a fix

Aliases change what is displayed; the underlying value is untouched. So an alias cannot repair a join key, cannot fix a duplicate caused by casing, and does not travel to another workbook.

If two spellings must genuinely become one value, use a Group, a calculated field, or fix it upstream — asserted in 10_tableau_prep.py, which shows the displayed values changing while the stored ones stay S and N.

⚠️ Tableau's "Pivot" means Power Query's "Unpivot"

Same operation, opposite name. Tableau's Pivot turns a column per month into one row per month — which Power Query calls Unpivot and pandas calls melt. Getting the vocabulary right per tool is worth a mark, and mixing them up loses one.

1. Extract filters        (what enters the .hyper at all)
2. Data source filters    (applied to every worksheet)
3. Context filters        (create a temporary subset)
4. Dimension filters      (the normal ones)
5. Measure filters        (after aggregation)
6. Table calculation filters  (applied LAST, hides without recomputing)

3.5 Calculated fields and LOD expressions

Calculated fields — the three kinds

⚠️ The same trap as Unit 2, in Tableau's words

WRONG:  AVG([Profit] / [Revenue])     -- row-level divide, then average
RIGHT:  SUM([Profit]) / SUM([Revenue])

Over the sample data those give 29.7619% and 27.3680% respectively — the same two numbers as Unit 2 §2.4, because it is the same mistake. Tableau makes it easier to commit, because dragging a ratio field onto a shelf aggregates it with AVG by default.

Aggregate, then divide.

{FIXED   [Region] : SUM([Revenue])}
{INCLUDE [Product] : SUM([Revenue])}
{EXCLUDE [Product] : SUM([Revenue])}

Pct of Region = SUM([Revenue]) / SUM({FIXED [Region] : SUM([Revenue])})

⚠️ FIXED ignores dimension filters — the classic surprise

FIXED is computed before dimension filters are applied, so filtering to one region does not change a FIXED [Region] result. To make a filter apply, promote it to a context filter — context filters run before LOD expressions. This is the same order-of-operations table as §3.4, and it is why that table is worth learning.

3.6 Joins and blending — and the trap between them

Joins

Tableau's joins are SQL joins, done at the data source. Database Management Systems' diagrams apply unchanged.

⚠️ The fan trap — the most valuable numeric example in this course

Join two tables that both have many rows per key, and your totals inflate. Nothing warns you; the numbers simply become wrong.

Take the nine sales rows and a targets table with two rows per store (one per quarter). Join them on store alone:

Revenue doubled — every sales row was duplicated once per target row. Targets more than tripled — each target row was duplicated once per that store's sales rows (Vijayawada has 4, Guntur 2, Hyderabad 3).

All four figures are asserted in 14_joins_blending.py.

The three fixes

3.7 Basic visualizations

⚠️ Scatter plots need a dimension on Detail

Drop two measures on the shelves and Tableau aggregates everything to one mark. You must put the identifying dimension (product, store, employee) on Detail to get one mark per thing. "My scatter plot has a single dot" is always this.

3.8 Storytelling and creating a Tableau story

Building a story

1. New Story, then drag a sheet or dashboard onto the first story point.

2. Caption each point — the caption is the narrative, and the default text is never good enough.

3. Use Duplicate to make the next point, then change one thing — a filter, a highlight, an annotation. Change one variable per step.

4. Add annotations to say what to look at.

5. Order the points as an argument, then Publish.

   Context   ->   Complication   ->   Cause   ->   Consequence   ->   Call to action
"Sales are    "But South fell    "Two large    "That is 8% of   "Reassign the
 ₹12.9 lakh"   11% in Q2"         accounts      annual revenue"   two accounts
                                  churned"                        this quarter"

3.9 Case study — HR analytics

Lab experiment 9, and the HR domain from Unit 1 §1.4.

The question: why do employees leave, and which groups are most at risk?

The data. One row per employee: emp_id, department, role, tenure_years, salary, last_rating, left (yes/no).

The measures:

That last one is why LOD matters here. {FIXED : ...} with no dimension computes over the whole table, giving the company-wide rate. Subtract it and each department shows its gap from the company average — the single most useful number on an HR attrition dashboard, and impossible without an LOD.

⚠️ The analytical trap in HR analytics

Attrition rate on small departments is unstable and will mislead. A team of 4 with one leaver shows 25%; a team of 200 with 40 leavers shows 20%. The small team looks worse and is not — one person's decision moved it 25 points.

Show headcount alongside every rate, and suppress rates below a minimum denominator. That is a Unit 5 dashboard-design point arriving early, and it is also Statistical Foundations for Data Science's sampling variability. Say so and you connect two courses.

Practice problems

Exam questions from this unit

Two marks

1. What does VizQL do?
2. Give the difference between .twb and .twbx.
3. What is the Marks card used for?
4. What do blue and green fields mean in Tableau?
5. Name the three LOD keywords.
6. What is a story point?

Five marks

1. Explain Tableau's architecture and its components.
2. Compare Live and Extract connections.
3. Explain the Tableau interface — shelves, marks card, views.
4. Explain the filter order of operations and why it matters.
5. Compare Tableau Public, Desktop and Reader.
6. How do you build a Tableau story? Describe the steps.

Ten marks

1. Distinguish joining from blending, illustrating the fan trap numerically.
2. Explain LOD expressions with a worked example.

3. Explain storytelling in Tableau and describe how you would build a story from an HR dataset.

4. Compare Tableau and Power BI across architecture, calculation language, data preparation and cost.

Mistakes that cost marks