Improving Code Reviewer Recommendation: Accuracy, Latency, Workload, and Bystanders
Bibliographic record
Abstract
Aim . The code review team at Meta is continuously improving the code review process. In this work, we report on three randomized controlled experimental trials to improve code reviewer recommendation. Method . To evaluate the recommenders, we conduct three A/B tests which are a type of randomized controlled experimental trial. The unit is either the code diff (Meta’s term for a pull-request) or all the diffs that an author creates during the experimental period. We set goal metrics, i.e., those we expect to improve, and guardrail metrics, those that we do not want to negatively impact, i.e., analogous to safety metrics in medical trials. We test the outcomes using a t -test, Wilcoxon test, or Fisher test depending on the type of data. Expt. 1 . We developed a new recommender, RevRecV2 , based on features that had been successfully used in the literature and that could be calculated with low latency. In an A/B test on 82k diffs in Spring 2022, we found that the new recommender was more accurate and had lower latency. The new recommender did not impact the amount of time a diff was under review. The results allowed us to roll-out the recommender in the Summer 2022 to all of Meta. Expt. 2 . Reviewer workload is not evenly distributed, our goal was to reduce the workload of top reviewers. Based on the literature and using historical data, we conducted backtests to determine the best measure of reviewer workload. We then ran an A/B test on 28k diff authors in Winter 2023 on a workload-balanced recommender, RevRecWL . Our A/B test led to mixed results. When a low workload reviewer had reasonable expertise, authors selected them, however, the top recommended low workload reviewer was often not selected. There was no impact on our guardrail metrics of the amount of time to perform a review. This workload-balancing replaced the recommender from the first experiment as the recommender in production at Meta. Expt. 3 . Engineers at Meta often select a team rather than an individual reviewer to review a diff. We suspected the bystander effect might be slowing down reviews of these diffs because no single individual was assigned the review. On diffs that only had a team assigned, we randomly selected one of the top three recommended reviewers to review the diff with BystanderRecRnd . We conducted an A/B test on 12.5k authors in Spring 2023 and found a large decrease in the amount of time it took for diffs to be reviewed. We did not find that reviewers rushed reviews. The results were strong enough to roll this recommender out to all diffs that only have a team assigned for review. Implications . Aside from the direct findings from our work, our findings suggest there can be a discrepancy between historical backtesting and A/B test experimental findings, and that more A/B tests are necessary to test recommenders in production. Outcome measures beyond accuracy are important. This is especially true in understanding how recommenders change a reviewer’s workload. We also see that the latency in displaying a recommendation can have a large impact on how often authors select recommendations making the reporting of latency an important metric for future work.
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 | 0.169 | 0.479 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.003 | 0.005 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.004 | 0.003 |
| Insufficient payload (model declined to judge) | 0.006 | 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".