DR.FIX: Automatically Fixing Data Races at Industry Scale
summary
Dr.Fix combines program analysis with LLM-driven generation (GPT40) and a vector database of 272 past bug–fix examples to auto-propose patches for Go data races, compiling and regression-testing each candidate with a feedback loop. Deployed at Uber across 100M+ lines, it produces fixes for 224 of 404 races (55%), sees 193 accepted by developers (86% of patches), and reduces median repair time from 11 to 3 days.
AI-written summary · about
chapters (16)
- 0:00Dr. Fix and data races
- 2:50Manual fixes challenges
- 3:52LLMs for auto-fixing races
- 4:59Bug embeddings and search
- 7:25Code locations for races
- 8:30Generating and validating fixes
- 9:57Developer review and deployment
- 10:20Impact at Uber scale
- 11:45Fixed race categories
- 14:15Limitations of Dr. Fix
- 15:24Ablation study effectiveness
- 16:33Summary of Dr. Fix
- 17:42Q&A: Matching bugs with keys
- 18:24Q&A: Code quality impact
- 19:47Q&A: Three-day fix breakdown
- 20:51Q&A: Developer feedback
related talks
22mData race detection
01h 04mDebugging data races
14mShow Me Why It's Correct: Saving 1/3 of Debugging Time in Program Repair with(…)
36mWhy programming is fiendishly difficult and what we can do about it
01h 00mSOAP - Keynote: Formal Reasoning and the Hacker Way
14mThe Power of Backports : Business Impact and Automation Strategies
28mLLOR: Automated Repair of OpenMP Programs
07mRace Conditions in Details