Build06 Standards & CI/CD · 3 of 3

Definition of done

A definition of done is only useful if a reviewer can check every line of it in under a minute without asking the author anything.

Stand 29.07.2026

#A definition of done that bites

What

The usual version — code reviewed, tests written, documentation updated — cannot be checked, so it is never enforced, so it is decoration. Each row below names the artefact and where the reviewer looks for it.

Definition of done for a warehouse pull request

Done means	Reviewer checks
Grain is declared	the table COMMENT contains a 'one row per …' sentence, in the DDL in this diff
The transform is importable	a test imports it directly, with no pipeline and no table; no dbutils, spark or widget read appears at module level anywhere outside an entry/ notebook
Business rules are asserted, not described	each rule touched by the diff maps to an EXPECTATIONS entry or a .sql assertion returning zero rows
Behaviour-neutral changes are proven neutral	a symmetric EXCEPT ALL result is linked in the pull request
The bundle still validates	bundle validate -o json is green for every target including prod, run with the same --var values the release uses — not only test
No literals that vary by environment	grep the diff for dwh_dev / dwh_prod; the catalog arrives as a variable
Identifiers pass the rule	^[a-z][a-z0-9_]*$ — the CI identifier step is green for object names; new COLUMN names are read by eye, because the grep does not cover them and no new identifier in the diff needs backticks
Ownership is real	the object is owned by a group; no individual appears as owner in the DDL
The re-run question is answered	the runbook says whether this job is safe to re-run blind, updated here if the answer changed
PII is classified	any new column carrying personal data has its governed tag applied in the DDL, not clicked in the UI

Ten rows. Each is binary, each is visible in the diff or in the build, and none of them requires the reviewer to trust the author. That is the whole design goal.

Why

An unenforceable definition of done fails in a specific and expensive way: the first three sprints look fine, then the first person leaves and it turns out nothing was written down. The grain lived in someone's head, the re-run question was answered in a chat thread, and the tags were clicked in the UI by a contractor whose access has since been revoked.

It also decides whether handover is possible. Every row above is something the team must be able to check without you in the room. A row that only works when the consultant reviews it is not a standard — it is a dependency.

Doing it

Put the ten rows in the pull-request templateThe reviewer then reads them at the moment they matter.
Automate every row that can be automated — identifiers, environment literals, bundle validateLeave the rest as explicit reviewer questions.
Add a row only when a real defect escaped that it would have caughtName that defect in the commit message.
Delete a row that has never once failed, and timebox reviewIf checking the list takes more than five minutes, the pull request is too big and that is the finding.

Embedding it

Have the team draft the list and argue you down on two rowsA list they negotiated is one they apply when you are gone.
Rotate the reviewer role weekly so no single person becomes the definition of done
Once a month, check a merged pull request against the list in front of everyoneFind one miss — there is always one — and add the automation rather than the reprimand.
Stop reviewing personally at the point they catch more than you doSay the date out loud; that is the handover milestone, not the final presentation.

Defaults

Ten rows or fewer, each binary, each visible in a diff or a build
Anything automatable is automated; the manual rows are the ones needing judgement
The list lives in the pull-request template in the repository, versioned like code
Rows are added in response to escaped defects and removed when they stop firing

Gotchas

Rows that cannot be checked — 'code is clean', 'documentation updated'They train reviewers to tick boxes without looking, which devalues the rows that do matter.
A twenty-row listReview becomes a chore, people tick everything, and the list now provides false assurance rather than none.
Enforcing it only on the consultant's pull requestsThe standard has to bind everyone from week one or it reads as an outsider's rule and dies at handover.
No written exception pathThere will be an emergency fix; without one carrying an expiry date, the emergency becomes the precedent.