Context
Learned indexes treat database indexes as predictive models that estimate where a key should appear in sorted data. ALEX is an adaptive learned range index that uses a hierarchy of regression models to support point lookups, range queries, inserts, updates, and deletes, but its original design relies on linear models and can degrade when local key distributions are highly nonlinear. This project explored that limitation by modifying ALEX toward polynomial, specifically quadratic, model-based indexing.
What I contributed
I forked and modified the ALEX C++ codebase to prototype a polynomial learned-index variant which was then implemented in DuckDB. My work focused on replacing the original linear model assumption inside the node structure with a polynomial model, preparing the code for build and test execution, and using the existing ALEX benchmark structure as the basis for evaluating whether higher-order models could improve prediction behavior on nonlinear key distributions.