How implementing ML ended up with an if-else
summary
Project predicts whether an OpenShift cluster upgrade will break, using OpenShift cluster streams and months of model experiments (decision trees, classical ML and even LLMs) but finds a simple rule beats or matches them given the data. Dataset contains less than 1,000 failures, is highly imbalanced, and the product optimizes for minimizing false positives via a precision-weighted F-score (alpha 0.55). Launched as a preview rule‑based feature (about three lines of code) that runs cheaply in production while collecting more signals.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:23Agenda
- 1:33Context
- 2:36Problem Statement
- 7:48Typical Problem
- 8:41Data Set
- 9:46Moral Challenges
- 10:31AI ML
- 10:59Metric choice
- 11:20Prediction
- 11:44False positives
- 13:57Fscore
- 14:59Baseline
- 17:05Iterating
- 18:00Issues
- 19:24Theories
- 24:05Qualification challenges
- 28:11Scalability
- 29:47Questions
related talks
27mPredicting Faulty Validations in Cluster Issue Detection: A ML Approach
42mA/B testing with real-world data to evaluate ML model updates
30mContainer workflows for data science and ML
10mOptimizing Machine Learning (ML Ops) in Financial Services with Kubernetes and Cloud Technologies.
25mAutomating Operations with ML
34mML-Bound Anomaly Detection: Experiences Beyond Commodity Observability
46mPanel: OpML
18mBeyond Inference: Bringing ML into Production