Toward the optimization of peer review effort in DevOps: an empirical investigation using merge request traces

Abstract

Peer review is a central stage of modern DevOps workflows, where software changes are examined before integration through mechanisms such as merge requests (MRs). Although prior research has studied peer review from several angles, review effort is still commonly approximated by temporal measures that conflate review work with waiting time, reviewer availability, and organizational urgency. This thesis studies peer review effort as a software process phenomenon rather than a purely temporal indicator, and investigates how it can be better characterized, analyzed, and optimized in DevOps environments. The thesis follows an empirical approach based on MR traces from two industrial partners (Kaloom and TELUS) and open-source GitLab projects. It establishes MR traces as an observability instrument that captures review activity together with environmental disruption, technological transitions, organizational priorities, and automation–human interactions. It introduces the amount of post-submission change as a time independent effort indicator, shows that it captures a distinct dimension of review work, and predicts it at MR creation time with an AUC of up to 0.88. It examines how review effort and quality vary across development, configuration, and documentation MRs, highlighting the specific profile of configuration changes and the amplifying effect of bug presence. It defines a taxonomy of workflow deviations, shows that they affect up to 37% of industrial MRs, detects them with up to 91% accuracy, and demonstrates their impact on effort analysis. It then extends the analysis upstream, by linking review effort to task estimation, and downstream, by relating review decisions to CI/CD outcomes and predicting safely self-approvable MRs with an AUC above 0.82. Finally, it presents RevMine, an LLM-assisted tool for reproducible DevOps-pipeline data collection and analysis across GitHub and GitLab, which integrates Kanban-style planning data, peer review data, and CI/CD delivery data into a single cross-stage workflow and thereby operationalizes the full upstream–review–downstream chain studied in the preceding chapters. Across these contributions, the thesis argues that peer review effort is a multi dimensional construct that no single indicator captures adequately, and that its optimization is a calibration problem rather than a minimization one: effort should be reduced where evidence shows it is unnecessary and preserved where it is under-invested. By connecting upstream planning, the review stage itself, and downstream delivery outcomes, the thesis provides empirical foundations, methodological artifacts, and tooling support to interpret and manage peer review effort in DevOps settings.

Date	8 Jul 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Francis Bordeleau (Supervisor)