SESSION
View the slides 
Moving Faster and Reducing Risk: Using LLMs in Release Deployment
This talk discusses the challenge of determining what should be released in large-scale software development, such as at Meta’s scale. To address this, we developed models to determine the risk of a pull request (diff) causing an outage (aka SEV). We trained the models on historical data and used different types of gating to predict the riskiness of an outgoing diff. The models were able to capture a significant percentage of SEVs while gating a relatively small percentage of risky diffs. We also compared different models, including logistic regression, BERT-based models, and generative LLMs, and found that the generative LLMs performed the best.
