All lessons

Module 6: Deliver BI people trust · Lesson 6.7 · about 40 min

Project 6: BI people can trust

Ship chain-dashboards: Power BI in git, an open-source twin, an idempotent morning email and reconciliation evidence a manager understands in a minute.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 6 project · about 40 minutes to read, a day or two to build · read first: Lessons 6.1 to 6.6. You’ll ship the BI layer on top of Projects 3 to 5: a Power BI report kept in git, an open-source twin, an automated morning email, and the reconciliation evidence that proves the numbers, all explained so a non-technical manager gets it in a minute.

A report is trusted only when every link holds at once: correct data, correct measures, fresh on time and clear to read, and the project's job is to prove each link with evidence.

1. What you are shipping

chain-dashboards, on the marts from Projects 3 to 5 (the restaurant group and the hotel group, all synthetic):

  • Power BI, as PBIP in git: a restaurant operations page, a hotel revenue page, a drill-through, 20 to 25 documented measures, RLS roles and the freshness banner.
  • An open-source twin in Metabase (or Superset or Evidence) on Postgres, with the same numbers.
  • A daily email from Dagster: HTML plus CSV, a partial-data banner, a sent record and a test mode.
  • Reconciliation: the R104 bridge and a daily control-total check that blocks the report.
files ─► Dagster (land, clean, load) ─► dbt marts ─► control-total check ─┬─► Power BI refresh ─► pages
                                                                         ├─► Metabase twin
                                                                         └─► 08:00 email (sent-guard)

2. The chain of trust

Correct data. The conservation test (Project 5) and the control-total check (Lesson 6.6) prove that what reached the marts is what the sources sent.
Correct measures. Five key measures are cross-checked against SQL on the warehouse, including a total row and a day with no sales (Lesson 6.2).
Fresh on time. The refresh is triggered by the load, and the "data as of" time comes from the audit table, with a banner for missing sources (Lessons 6.3 and 6.4).
Clear to read. One page per audience, KPIs with comparisons, sorted bars, and someone who has never seen it finds yesterday's weakest restaurant in ten seconds.
Multiply. At 99 %, 90 %, 97 % and 95 % the chance all four hold is about 82 %. Raise the 90 % link (measures) to 99 % and it climbs to about 90 %. Fix the weakest link first.
Maths and principle

Principle: a report’s credibility needs every link at once, so it behaves like a product of probabilities, and one weak link drags down the whole.

Equation:

P(trusted)≈P(data)⋅P(measures)⋅P(fresh)⋅P(clear)

Four links at 95 % each give 0.954≈81%.

Analogy: a chain holds only as well as its weakest link, and a string of fairy lights wired in series goes dark if any one bulb fails.

Where it breaks: the links are not fully independent (a late file hurts both freshness and correctness). The lesson still holds: find and fix the weakest link.

Try it: data 99 %, measures 90 %, fresh 97 %, clear 95 %. Which improvement to 99 % helps most? (Measures: the product rises from about 0.82 to about 0.90.)

The daily control-total check is the piece that keeps the “correct data” link high without anyone watching:

PBIP
The Power BI project format: model and report as text files in git.
Open-source twin
The same pages in Metabase, Superset or Evidence, proving the logic lives upstream.
Chain of trust
Correct data, correct measures, fresh on time, clear to read; all four must hold.
Design notes
A README section listing audiences, their questions, and what you left out and why.
Physical meaning
A late hotel file on a Monday shows up as a banner saying "H3 not received", not as a fake 20 % drop in occupancy. The general manager reads the banner, ignores H3 for the morning, and acts on the four hotels whose numbers are real.

“Publish to web is the easy way to show my report.” It creates a public link anyone can open, is often switched off by company administrators, and must never be used with real company data. Use screenshots, a short video and a synthetic-data twin instead.

3. Acceptance checks

# Requirement Proof in the repo
1 PBIP committed; every measure has a description and a display folder repo
2 Five measures cross-checked against SQL, including a total and a blank day docs/measure-tests.md
3 RLS roles tested with View as screenshots
4 “Data as of” from the audit table; missing-source banner a screenshot from a day with a missing source
5 Open-source twin with matching numbers, plus a costed tool comparison docs/bi-comparison.md
6 Email report with sent record and test mode a Dagster run and a screenshot of the received email
7 Reconciliation document, waterfall and a blocking control-total check docs/reconciliation/ and the check’s code
8 Design notes: audiences, questions, what was left out README

4. Making it easy for reviewers

Reviewers will not install Power BI. Give them, at the top of the README: screenshots (PNG), a 60 to 90 second video, the PBIP for anyone who wants to open it, and optionally a public Metabase or Evidence version with synthetic data only. Make sure no file contains credentials: connection details are parameters, and the data is synthetic.

5. Build checklist

  1. Tidy the pages: consistent colours, a phone layout for the KPI page, every visual answering a listed question.
  2. Final run: generate fresh data, run Dagster from end to end, let it trigger the refresh and the email, and record the times in the README.
  3. Break it on purpose: delete one hotel file and record what the banner, the email and the control-total check do.
  4. Screenshots and video, then the README: summary, the chain-of-trust evidence (one line per link with a link to its proof), design notes, how to run it, next steps.

Your story (fill in your own numbers): “I delivered the same restaurant and hotel marts through Power BI and Metabase. The Power BI model is in git as PBIP with [N] documented measures cross-checked against SQL, row-level security by region and brand, and a ‘data as of’ label from the pipeline’s audit table, so a late hotel file shows as a banner instead of a fake drop in occupancy. A store challenged one month’s sales; my bridge explained the gap as tax, business-day timing and refunds, with a $[x] residual traced to [y] quarantined receipts. That is why there is now a daily control-total check that blocks the report.”

What done looks like: a manager who has never seen the project watches the video and can say what the dashboard is for; a reviewer can follow each acceptance line to its proof; breaking a source produces a visible, honest message instead of a wrong number. Commit the final state as feat: project 6 chain-dashboards.

From here, Module 7 runs all of this in production. When your data outgrows one machine, the Data at Scale course picks up with Spark, BigQuery and Kafka.

  • The project proves four links of trust: correct data, correct measures, fresh on time, clear to read.
  • Trust behaves like a product, so improve the weakest link first; measures are usually it.
  • Show reviewers screenshots, a short video, the PBIP and an optional synthetic-data twin; never publish real data.
  • A broken source must produce a visible banner and a blocked check, never a silently wrong number.

Check yourself

1Reviewers do not have Power BI. What is the best way to show your report?
2Which link most often breaks trust in BI?
3Four links of trust at 99 %, 90 %, 97 % and 95 %. Which one should you improve first?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. All company data in the examples is synthetic. Tool versions and prices change: check the official docs before you rely on them.