Engineering Practice
Measure Developer Productivity Without Ranking People by Commit Counts
A filled team measurement agreement defines system-improvement signals, ownership and reporting safeguards. Actual arithmetic on synthetic windows shows commits rising while waits and reported focus worsen, leading to a bounded routing trial without personal rankings or causal claims.
Measure developer productivity to find and improve constraints in the team's work system. Combine delivery evidence, quality, collaboration and developers' experience around a stated question. Commit counts can describe activity in a defined repository, but they cannot establish who contributes most or whether customers receive better software. Avoid turning a team improvement review into a personal league table.
This LandOffer.ai guide includes a completed measurement agreement and a two-window assessment using entirely synthetic data. The calculations were actually executed locally; the team, discussions, records and responses are fictional. No employee monitoring, repository extraction, real survey, performance evaluation or organizational experiment was performed. The case demonstrates a reasoned improvement decision, without claiming that a productivity intervention worked.

Start with a team question that measurement can answer
The Microsoft Research SPACE publication abstract says developer productivity extends beyond individual activity and cannot be measured by one metric or dimension. The abstract supports that general principle; it does not by itself supply a complete measurement procedure or evidence about your organization.
In her first-person explanation of SPACE, co-author Nicole Forsgren describes satisfaction, performance, activity, communication and efficiency. She explains why commits, code volume and completed pull requests miss contributions such as collaboration, documentation and helping colleagues. Ask which parts of the team’s work each signal misses. Do not turn the dimensions into five equally weighted employee scores.
“Who committed the most?” and “Why are reviewed changes waiting longer?” lead to very different records. The second question points toward a system boundary, a delay and a possible decision. Define what the team needs to learn before choosing which tool can export the most data.
For example, choose one application and a consistent review process. Identify when a change becomes ready for review, what counts as a substantive response, and whether waiting time uses elapsed or working hours. Report the open queue beside waits that have ended, so a shorter median cannot conceal unresolved older requests. A completed-item median alone can improve while the oldest blocked work remains unfinished.
Completed measurement agreement: one service and one improvement purpose
The fictional HarborElm team maintains a document-preview service. Its 12 engineers and engineering manager Mira agree to investigate whether the review process is interrupting delivery and concentration. The measurement purpose is to choose a process improvement, not determine compensation, promotion, discipline or individual output targets.
The agreement covers two adjacent 14-day windows: October 5–18 and October 19–November 1, 2026. It keeps the service, repository scope and definitions constant. Mira records that the second window included a planned parser migration and two production incidents. Those circumstances may matter; their contribution to any observed change has not been estimated.
Here is the completed signal portfolio. It is the team's invented measurement design, not a validated SPACE questionnaire or a complete implementation of DORA.
| Signal and purpose | Definition and collection boundary | Interpretation limit and owner |
|---|---|---|
| Commit activity, as context | Count non-bot commits reachable on the service's default branch during each window, using the same inclusion rule | Squashing and commit size affect the count. Analyst Dev owns the definition; no author breakdown or productivity target is produced. |
| Review waiting, for coordination and flow | Elapsed hours from marked ready to first substantive human review, for items receiving that first review within the window; exclude draft time and automated comments. A completed wait does not mean the pull request merged. | Completion selection can hide older work. Dev also reports open reviews older than 24 elapsed hours at each window end. |
| Deployment volume, for delivery context | Count production deployments of this service in the window, using the same deployment identifier rule | Volume alone does not measure useful customer outcomes, scope or engineering effort. Release lead Sal owns the record. |
| Intervention after deployment, as a quality guardrail | Count deployments requiring immediate intervention, divided by all production deployments in that window | Small denominators make the ratio unstable; inspect the actual events. Sal verifies inclusion against the stated definition. |
| Perceived focus, for experience | Voluntary question: “During this window I had enough uninterrupted time for my planned engineering work,” scored 1–5; report responses 4–5 as favorable | This is a local, unvalidated question. Response composition can differ between windows; Mira cannot identify or match respondents. |
| Collaboration context | A facilitated discussion of review handoffs, question routing and work that helped others | Use agreed summaries of process issues, not named complaints, message counts or a helping-people leaderboard. Mira facilitates. |
The team names Dev as the designated analyst for raw survey responses. Mira receives the permitted aggregate and agreed process summary, without individual responses or respondent identities. The fictional reporting rule requires at least eight responses before showing an aggregate; it is a local safeguard, not proof of anonymity or a universal safe threshold. No demographic slices, timestamps or verbatim identifiable comments enter the review packet.
Explain the intended use before collecting anything in a real team. Apply the actual privacy, access and retention requirements, and check whether a small group could still be identifiable. Do not promise anonymous feedback while retaining names or combining it with information that reveals a respondent. Collect only what the improvement question needs.
Activity rises while several other signals worsen
The completed fictional review uses these inputs and calculations. Each review-wait sample is an authored observation whose wait ended; the 8-value and 6-value arrays are saved with the executed arithmetic trace. They are not pull-request lifecycle or merge-duration measurements. There are no underlying real people or repositories behind them.
| Signal | October 5–18 | October 19–November 1 | What the comparison supports |
|---|---|---|---|
| Included commits | 40 | 72 | Activity count rises by 32, or 80%; this is not an 80% productivity gain. |
| Completed review waits, elapsed hours | 1, 2, 2, 3, 3, 4, 4, 5 | 2, 4, 5, 6, 8, 9 | Median rises from 3 to 5.5 hours; samples contain 8 and 6 completed observations respectively. |
| Open reviews older than 24 elapsed hours | 3 | 7 | More aged open work is present at the snapshot; it is not included in either completed-item median. |
| Production deployments | 8 | 6 | Recorded deployment volume decreases, without establishing how much user value each deployment contained. |
| Deployments requiring immediate intervention | 1 of 8, or 12.5% | 2 of 6, approximately 33.3% | The observed ratio increases; the small sample cannot establish a stable future failure probability. |
| Focus question | 6 favorable of 10 responses: 60% | 2 favorable of 8 responses: 25% | The favorable share falls among respondents; these are not paired responses from the same identified people. |
Both survey windows have 12 eligible people. Ten responses and eight responses correspond to response rates of approximately 83.3% and 66.7%. Keep those denominators next to the favorable shares. The unanswered questionnaires are not unfavorable responses, and a change in response rate can change whose experience the aggregate reflects.
The local Python calculation completed 24 checks of dates, window adjacency, denominators, scale values, medians and ratios. It confirms arithmetic on authored inputs, not the correctness of a real extraction or the validity of the survey. It does not provide a significance test, confidence interval or causal estimate.
The review therefore rejects “productivity increased 80%.” It also rejects the opposite leap, “the engineers became less productive.” The packet shows more recorded commit activity alongside longer review waits, more aged work and less favorable reported focus. That combination warrants investigating the work system and the changed context.
Discuss the observations without assigning blame
Mira opens the November 2 fictional review with this readback:
We have more commits, but that does not tell us who accomplished more. Completed review waits and the aged queue both worsened. Focus responses are less favorable, with fewer respondents. We need to understand the review handoff while preserving quality; we have not established that the parser migration, incidents or any person caused these differences.
The agreed discussion summary records that review requests sometimes waited without a named route when the usual domain reviewer was handling an incident. Pairing and explaining parser behavior also occupied time without consistently creating commits. These are fictional participants' reported process observations, not independently timed explanations for every delay.
Dev checks the metric definitions before proposing a fix. Both windows used the same ready marker and excluded automated comments; there was no stated collection-rule change. That reduces one specific comparability concern but does not prove identical work complexity or staffing availability. The team does not divide commits by headcount to manufacture an individual efficiency score.
The current DORA guide describes five software delivery performance metrics, including deployment rework rate alongside the other measures. It recommends application or service context and improvement over competition; it warns against single-metric goals and misleading comparisons across dissimilar applications. This portfolio borrows a narrowly defined deployment guardrail, not a claim to have measured overall developer productivity with DORA.
The team has not measured task success for document-preview users or the value of the parser migration. Sal records that gap explicitly. A smaller number of deployments might contain a valuable change, and a larger number might mostly repair avoidable problems. The packet cannot decide between those explanations from counts alone.

Complete the improvement decision and its limits
The November 2 decision assigns one bounded change: try an explicit domain-review route for the next two weeks. Sal accepts responsibility for publishing the current review owner and an available backup before the trial begins; the backup must accept capacity rather than be assigned silently. Mira accepts checking coverage at the team's existing planning meeting. The trial changes the stated routing process; it creates no instant-review guarantee or new on-call obligation. Coverage must fit accepted capacity.
Dev will use the same completed-wait definition and aged-open snapshot, and repeat the optional focus question with the same reporting safeguards. Sal will continue the intervention guardrail. The team will not skip required reviews, shrink tests or split commits solely to improve a count. If the backup is unavailable, the request stays explicitly blocked and Mira resolves priority instead of treating an unaccepted tag as coverage.
The review is scheduled for November 16. Its decision question is whether the routing change was usable and whether the combined evidence supports continuing, revising or stopping it. There is no promised percentage improvement, and the chosen two-week period is a practical local trial length, not a statistically justified detection window.
The completed decision record contains the baseline, observed concern, accepted responsibilities, unchanged measurement rules, quality guardrail and review date. It contains no post-trial results. No routing change, meeting invitation, employee survey or repository operation was actually carried out for this article; only the synthetic arithmetic was executed.
Further workplace guides are available from LandOffer.ai. Keep team signals attached to an improvement question, include the experience and work that activity counters omit, and review the resulting change with the people doing the work. At the review, record what was actually tried, what remains uncertain and whether the team accepts continuing the change; do not report the planned trial as a verified improvement.
Sources and Further Reading
- Microsoft Research: The SPACE of Developer Productivity — original-author publication abstract, February 2021; abstract read, not a claim to have accessed the blocked full ACM article.
- Nicole Forsgren, Microsoft Azure: Navigating the SPACE Between Productivity and Developer Happiness — first-person framework explanation, May 11, 2023; checked October 28, 2026.
- DORA: Software Delivery Performance Metrics — current five-metric model, service context and measurement pitfalls; updated January 5, 2026; checked October 28, 2026.


