Compute Engine Testing with Synthetic Data Generation
summary
Meta builds a Synthetic Query Bank that runs privacy-safe, production-like tests against Presto (classic, Velox, Presto on Spark) to detect correctness, performance and reliability regressions. It generates tables with a two-step DP synthetic data pipeline: a differentially private synthetic data GAN-like model for features and a second-order histogram DP-PUI algorithm for user-identifying columns, flattens arrays/maps/timestamps, and stores stable baseline checksums. Results: >60% of queries match output row counts exactly, enabling faster, reusable PR-time end-to-end tests.
AI-written summary · about
chapters (12)
- 0:00Compute Engine testing intro
- 1:26Synthetic Query Bank testing
- 2:51Privacy-compliant testing
- 4:18Faster signal with less data
- 5:43Correctness regression detection
- 6:24Performance regression baselines
- 7:50Synthetic Query Bank use cases
- 8:10End-to-end query bank workflow
- 10:19Differentially private data gen
- 12:08DP UI column preservation
- 14:16Query bank evaluation
- 15:43Summary
related talks
16mGenerating High-Quality Tabular Synthetic Data at Scale
01h 00mThe Synthetic Data Advantage: How It Helps You to Get Both AI Privacy & Explainability
13mDPSynth: From Research to Production—Engineering Differentially Private Synthetic...
01h 14mStatistical Inference and Privacy, Part III
28mInsights, Not Secrets: How Synthetic Data Empowers Privacy
14mToward Provably Private Insights into AI Use
21mCrafting Flight Simulations: The Art of Synthetic Data Generation
20mSynthetic Data for Privacy, AI Fairness and Democratizing Data Access