Free checklist · Power BI · 20 checks

Is your Power BI model telling the truth?

Twenty checks I run on every report I review, in four groups: the numbers, the model, the speed and who can see what. Most take minutes. Any one that fails is worth knowing about before an auditor or a board finds it.

milestonebi.com/checklist · Omar Williams, Milestone BI

Get the PDF →

01 / Numbers

The numbers

1 Ratios are summed, then divided once.
Averaging a column of ratios gives a different answer: 22.7% instead of 16.2% for EU bank capital. How it happens.
2 Every rate has the right denominator.
A fraud rate divided by every transaction, when only two thirds were ever checked, comes out a third too low. See the report.
3 Key measures reconcile to source totals.
The check is written down, not eyeballed, so the next person can run it again.
4 Totals mean what they say.
A grand total that isn’t the sum of its rows is either deliberate and labelled, or a bug.
5 Year-on-year compares like with like.
The same months against the same months, not a part year against a full one.

02 / Model

The model

6 Star schema, with one documented grain per fact table.
Anyone can say what one row of each fact table means without opening the data.
7 Keys really are keys.
The “one” side of every relationship is unique. A product key that isn’t quietly duplicates rows. See the build.
8 Relationships filter one way by default.
Every bidirectional filter has a stated reason.
9 There’s one date table, marked as a date table.
And auto date/time is switched off, so no hidden date tables are inflating the model.
10 Every figure on a page is an explicit measure, with a description.
No implicit sums of columns dragged onto a visual.

03 / Speed

The speed

11 Unused columns are removed.
One flat table carrying everything used six times the memory of a star schema. See the measurements.
12 The slowest visuals are known.
Measured in Performance Analyzer, not guessed.
13 Heavy measures have been profiled.
None iterates a large table where a simple aggregate would do.
14 Model size, and what drives it, is known.
Measured with VertiPaq Analyzer, so the biggest columns are a decision rather than an accident.

04 / Access & governance

Who can see what

15 Row-level security roles exist and have been tested.
Tested with “View as” for each role, not assumed from the role definitions.
16 Workspace admins are groups, not individuals, and there are few of them.
Milestone Pulse flags this from a workspace export in seconds.
17 No orphaned workspaces or datasets are left behind.
Every workspace has an owner and something in it that someone still uses.
18 Refresh credentials don’t belong to a person who could leave.
Gateways and data source credentials sit with a service account or a group.
19 The model is in source control as a PBIP project.
Changes can be reviewed, compared and rolled back.
20 Someone other than the author could explain every measure.
If only one person can, the model leaves when they do.

Want the checks run for you? A fixed-price health check, £495, three working days: milestonebi.com/power-bi-review

Take it with you

Get the checklist as a PDF.

One page per group, with space to note what you find. Leave your email and the download opens straight away.

Fixed price · three working days

Rather have the checks run for you?

See the £495 health check →