Module 6: Deliver BI people trust · Lesson 6.5 · about 35 min
Open-source BI without Microsoft licences
Run Metabase and Superset on your marts, place Evidence, Lightdash and Grafana, and pick a BI tool for a medium company with a costed break-even.
Module 6 · about 35 minutes, plus the lab · read first: 5.6, 6.3. You’ll be able to run Metabase and Superset on your marts with Docker, rebuild a Power BI page in Metabase, place Evidence, Lightdash, Grafana and Streamlit, and recommend a BI tool for a medium company with numbers behind it.
1. Why this matters
Power BI is cheap per user, but the cost grows with every viewer: 200 restaurant managers is 200 licences under most plans. Some companies run Google Workspace or no office suite at all. Open-source BI tools are mature, free to self-host, and in many cases good enough. A senior engineer can recommend either, with a costed comparison.
Choosing a BI tool is like choosing a shop front for a bakery. The ovens, recipes and quality checks (the pipeline, dbt, tests) stay the same whichever window you put the bread in. If the recipes lived inside the window display, changing the window would mean rewriting the recipes.
2. The options
| Tool | Style | Strong at | Weaker at |
|---|---|---|---|
| Metabase | Point-and-click questions plus SQL, dashboards, subscriptions | Fastest for business users; email and chat subscriptions; embedding | Complex modelling; very large semantic layers |
| Apache Superset | SQL Lab plus a chart builder with many chart types | Many users, role-based access, row-level security, rich charts | Harder to set up and to author in |
| Evidence | Reports as code: Markdown plus SQL, built as a static site | Lives in git, looks polished, cheap to host | Not for self-service exploring |
| Lightdash | BI on top of your dbt project | Uses dbt YAML as the semantic layer | Needs dbt |
| Grafana | Time-series and operations dashboards | Pipeline monitoring and alerts | Not a business BI tool |
| Streamlit or Dash | Python apps | Custom tools: what-if pricing, schedule builders | You maintain the code |
Pricing and features of the paid editions and cloud versions change; check each project’s site before you write numbers into a recommendation.
3. Metabase in ten minutes
# docker-compose.metabase.yml
services:
mb-db:
image: postgres:16
environment: { POSTGRES_USER: mb, POSTGRES_PASSWORD: mb, POSTGRES_DB: metabase }
volumes: ["mb-db:/var/lib/postgresql/data"]
metabase:
image: metabase/metabase:latest # pin an exact version for real use
ports: ["3001:3000"]
environment:
MB_DB_TYPE: postgres
MB_DB_DBNAME: metabase
MB_DB_PORT: 5432
MB_DB_USER: mb
MB_DB_PASS: mb # local trial only; use secrets in real use
MB_DB_HOST: mb-db
depends_on: [mb-db]
volumes:
mb-db:
docker compose -f docker-compose.metabase.yml up -d. Metabase keeps its own questions, dashboards and users in the mb-db Postgres, not in a file inside the container.report_reader login, limited to the mart schema.dim_date already has. No logic is rewritten; it is all in the mart.Then add the “data as of” card from mart.freshness (Lesson 6.3). Because the freshness logic lives in the warehouse, it took one query, not a rebuild.
4. Superset in brief
Use the official Docker Compose setup from the Superset repository for local trials; real deployments use its Helm chart or a managed provider. Connect with SQLAlchemy URLs (postgresql+psycopg2://..., mssql+pyodbc://...). Define datasets (a table or a SQL query), then charts, then dashboards. Row-level security is set per role and dataset.
Maths and principle
Principle: licence cost grows with users; self-hosting cost is mostly fixed (servers and people’s time). The cheaper option flips at some number of users.
Equation: with price per user , server cost , hours of upkeep at wage , and a small extra cost per user,
Analogy: renting a car per day versus owning one. Owning has fixed costs every month; renting costs more per trip. Use it often enough and owning wins.
Where it breaks: hidden costs on both sides: sign-in integration, training, support, and the analyst’s time building in a less familiar tool. Add them as hours before you trust the break-even.
Try it: licence $14 per user per month; a $60 server plus 4 hours a month at $50; extra cost per user about $0. Break-even? ((60 + 200) ÷ 14 ≈ 18.6, so from 19 users self-hosting is cheaper on paper.)
5. Choosing
- Microsoft 365 company, people who live in Excel: Power BI. Integration with Teams, Excel and Entra ID is worth the licence.
- No Microsoft, many viewers, few authors: Metabase. Subscriptions and simple questions cover most needs.
- A data team that wants many chart types, SQL first, strict access control: Superset.
- Reports published as code, daily or monthly: Evidence.
- A custom tool (what-if pricing, a schedule builder): Streamlit next to whichever BI tool.
Whatever you pick, keep one set of metric definitions feeding it (Lesson 5.6). Moving a company from Power BI to Metabase then changes only the presentation layer.
- Self-hosting
- Running the open-source tool on your own server or cloud machine instead of paying per user.
- Application database
- The database where a BI tool keeps its own questions, dashboards and users.
- Presentation layer
- The BI tool on top; replaceable when the logic sits in dbt and the warehouse.
- Break-even users
- The number of users above which self-hosting costs less than licences.
- Physical meaning
- A 120-site deli chain with no Microsoft licences gets its managers' morning dashboard from a self-hosted Metabase on a small cloud server. Because the metrics live in dbt, the same numbers appear in the analysts' notebooks and in the email, and moving to another tool later would not change them.
“Open source is free.” The licence is free; the server, upgrades, backups, sign-in setup and someone’s time are not. Count them as hours, at a real wage, before you compare with a per-user licence.
6. Try it
- Load your marts into Postgres (the Module 4 warehouse resource with Postgres as the target).
- Run Metabase with a Postgres application database. Rebuild the restaurant operations page: KPIs with comparisons, trend, sorted restaurant bars, channel mix, and the “data as of” card.
- Set up a dashboard subscription to your own email through a test SMTP account.
- Run Superset locally; build the hotel pace chart and set one row-level security rule.
- Build the same restaurant page in Evidence and preview it locally.
- Write
docs/bi-comparison.md: time to build each page, what was harder, what each did better, and cost at 20, 100 and 300 users using the formula above with prices you look up on the vendors’ sites today (write the date you checked).
What to check: Metabase restarts without losing dashboards; the numbers on all three tools match Power BI for the same day; your comparison shows its sources and the date you checked prices. Commit as feat(bi): Metabase, Superset and Evidence versions with a costed comparison.
- Metabase suits many viewers and few authors; Superset suits SQL-first teams; Evidence publishes reports as code; Lightdash reads dbt; Grafana is for operations.
- Keep a BI tool’s own application database in a durable database with backups.
- Licence cost grows with users; self-hosting is mostly fixed cost. Work out the break-even, then add the hidden costs.
- Business logic in dbt makes the BI tool replaceable.
Check yourself
Sources
- Metabase documentation: running Metabase on Docker and the application database https://www.metabase.com/docs/latest/
- Apache Superset documentation https://superset.apache.org/docs/intro
- Evidence documentation https://docs.evidence.dev/
- Reis, J. and Housley, M. (2022), Fundamentals of Data Engineering, O'Reilly, Chapter 4 (choosing technologies; total cost of ownership)
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.