Kepler: Robust Machine Learning Methods to Reduce SQL Query Latency in Databases

What is it about?

Kepler is an ML-based optimizer for parameterized queries, tackling three main objectives: (1) Query Optimization: make queries execute faster (2) Parametric Query Optimization: reduce query planning time; fast model inference (3) Robustness: with high probability, don’t be worse than existing query optimizer, uniformly across the entire workload.

Featured Image

Photo by Pawel Czerwinski on Unsplash

Photo by Pawel Czerwinski on Unsplash

Why is it important?

We emphasize the robustness and deploy-ability angles of the machine learning + query optimization problem because it's hard to get adoption if your system can be suddenly worse than the existing optimizer! Our open source repository includes a Kepler sample deployment in Postgres to demonstrate end-to-end integration and performance. We include a research data set with ~14 years of execution data to enabled others to further the ML aspects of this work without the cost of data generation.