Research
Machine Learning for Systems
Learned components inside data systems — query optimization, indexing, caching and configuration tuning.
Database engines are full of decisions made by hand-written heuristics: which join order to pick, what to index, what to keep in memory, how to set hundreds of configuration knobs. Each of those is a prediction problem with a measurable cost, which makes them natural targets for learning.
Our interest is in learned components that survive production: models that adapt as workloads drift, that degrade gracefully when their predictions are wrong, and whose training cost is small enough to be worth paying. We study learned query optimization and cost estimation, workload-aware indexing and caching, and automatic configuration tuning.