Choose measures that can change a decision
Commits, lines of code, merge requests, pipeline runs, and accepted suggestions count activity. They don't show whether users received value, delivery became easier, or risk fell. When an activity count becomes a target, people can improve the number while quality, teamwork, and maintainability get worse.
Choose measures that help someone make a decision. DORA metrics can show delivery speed and stability. GitLab's DORA reporting can bring those measures together.1 Value Stream Analytics can point to a stage where completed work spends time. Quality, security, user results, and developer feedback add context that a delivery measure can't provide. None of them explains the cause by itself.
A chart can also encourage action before the question is clear. Higher deployment frequency might come from smaller releases or from duplicate deployment records. A shorter review stage might mean feedback arrives sooner or that review moved outside GitLab. Decide what you need to learn before choosing the number.
Connect a result to its costs and constraints
Start with the decision people face. What changed, why does it need attention now, and who experiences the problem? Describe a better result and what would make a proposed change unsuccessful. Faster releases may help customers receive fixes sooner, but won't be an improvement if they increase interruptions or make the release process harder to support.
Keep enough context with a measure to explain how you'll use it:
| Field | What to record |
|---|---|
| Decision | The choice this information can change and the person who can make it. |
| Scope | Products, services, projects, work types, and time period included. |
| Starting point | Current result, capture date, source links, and known gaps. |
| Main result | The change expected if the improvement helps. |
| Safety checks | Quality, security, stability, cost, or developer feedback reviewed beside it. |
| Review | Owner, review point, and the actions different results could trigger. |
If the team can't name a decision or a possible response, wait to add the measure. Keep collection effort tied to action, and retire a report when it no longer helps.
Choose a small set that describes different parts of the system. For example:
- a user or business result, such as a requested change reaching users or a service meeting its goal;
- a flow measure, such as stage time or lead time for changes;
- a quality, security, or stability check, such as failed changes, restoration time, escaped defects, or unresolved risk;
- developer feedback about waiting, setup, interruptions, confidence, or effort.
Review the measures together so the tradeoffs remain visible. A gain in speed shouldn't hide a rise in failed changes or developer effort. Use GitLab records to show what happened and interviews or a focused survey to help explain it. A single combined score would make those changes harder to distinguish and could reward a result the team wouldn't choose knowingly.
When a trend changes, open the issues, merge requests, pipelines, deployments, or incidents behind it. Separate normal work from urgent fixes, large migrations, and work that follows another release path. Ask what changed in workload, staffing, policy, or tooling, and compare the views of the people doing and supporting the work. The trend tells you where to investigate; examples and people’s accounts help explain the cause.
For example, suppose Value Stream Analytics shows a long Review stage. Sample merge requests and separate time waiting for a reviewer from time spent discussing the change. Talk with developers and reviewers. The cause might be unclear ownership, large changes, limited reviewer capacity, late automated checks, or a needed specialist review. Choose a trial that addresses the cause. Use review-stage time as the main result only for the affected work, and check follow-up fixes, production incidents, and developer feedback beside it. Keep the change only when the whole result improves.
Interpret change in context
The four DORA metrics have different source records. Deployment frequency needs successful production deployments. Lead time for changes also needs the merge request relationship. Time to restore service and change failure rate need GitLab incidents. Check how GitLab calculates each metric before using it as a safety check.
Comparison needs context. A service deployed many times a day and an application released on a fixed schedule don't have the same work pattern. A group result can hide one project with missing records or another project that dominates the total. Begin with a team's own trend; compare projects only when the work, definitions, and recording practice are similar enough to answer the question.
Aggregation changes the meaning of a chart. Deployment frequency may appear as a daily average, while APIs and other views can return interval counts. Several duration measures use medians, which describe typical completed work but can hide a small number of very slow items. Record the view, interval, and unit with the starting point.
Developer feedback needs a response path. Explain how answers will be used, keep the discussion about the delivery system rather than individual performance, and tell participants what changed. Collect feedback only as often as the owners can review and act on it.
Measures change behavior. Ask how someone could improve each number without improving the real result. Keep commits, merge requests, accepted suggestions, and other activity counts out of individual targets. Review measures with the team and change or remove them when they no longer help the team decide what to change.
Questions to ask
- What changed, why does it need attention now, and who cares about the result?
- What decision can this measure change, and who can make that decision?
- What work, period, source records, and known gaps define the starting point?
- Which quality, security, stability, cost, and experience checks belong beside the main result?
- What points to a system bottleneck rather than one person's output?
- Which work follows a different path and should be reviewed separately?
- What changed in definitions, workload, staffing, policy, or tools?
- How could someone improve the number without improving the result?
- What action did the last review cause the team to take?
- When should the measure be changed or retired?
Further Reading
Notes
-
GitLab's DORA reporting is available in Ultimate. The practice of measuring delivery outcomes doesn't depend on that feature; teams can use data from their existing delivery and incident tools. ↩
- Choosing measures that guide improvement shows how to compare a trial with its starting point and decide what mixed results mean for the next change.
- DORA metrics defines GitLab's metric inputs, calculations, assumptions, units, and reporting views.
- Value stream analytics defines stage measurement, item inclusion, lifecycle metrics, and work-item views.