Towards Sustainable AI for Continuous Integration Quality Gates
Bibliographic record
Abstract
Continuous Integration (CI) is the concept of continuously incorporating new code changes in the project's main codebase. To ensure the quality of the integrated changes, two types of quality gates are typically considered by most software development companies: automated gates, leveraging the build process, which invokes compilation commands and runs tests, and manual gates, based on peer reviews. With projects scaling up in size, the cost required for maintaining CI quality gates can be considerable (in terms of computational/manual effort). To assist or improve CI quality gates, the Software Engineering (SE) community has proposed a wide range of software analytics models. However, most of prior work focuses on the initial deployment of such models, without considering the different factors involved in their long-term sustainability: automated gates require smart retraining processes to maintain performance, while manual gates require synergy with reviewers’ workflows. In this thesis, we study sustainable AI for CI quality gates. For automated gates, we explore how to improve the learning process of software analytics models using (1) dynamic training scheduling strategies, where we evaluate different ways to dynamically adapt the training time for Retraining-From-Scratch (RFS) setups, and (2) Lifelong Learning (LL) training setups, where we change the fundamental learning algorithms toward an incremental setup. For the manual CI quality gates, we perform user studies to evaluate the impact on reviewers' productivity of (3) hotspot-based file-ordering, where problematic changes are shown first to the reviewers, and of (4) reviewers’ interaction with generated review comments. We were able to optimize the trade-off between model performance and the computational effort of model retraining, with our best heuristics recommending retraining only once every 5-6 weeks without losing performance to weekly retrained models. Yet, LL substantially reduces computational effort even further by 2-40x compared to RFS. Through large-scale industrial user studies, we also observe that reviewers benefit from ML-based assistance, as file-reordering led to more (+23%) and better (precision +13%, recall +8%) review comments compared to the default alpha-numeric file-ordering, while generative AI-based review comment generation obtains promising results regarding acceptance (8.1% and 7.2%) and appreciation (23% and 28.3%).
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.012 | 0.055 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.004 |
| Scholarly communication | 0.006 | 0.011 |
| Open science | 0.004 | 0.009 |
| Research integrity | 0.003 | 0.006 |
| Insufficient payload (model declined to judge) | 0.004 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".