Fast Hypothetical Updates Evaluation
Venue: ProvenanceWeek 2025
Abstract
We demonstrate FaDE, a DuckDB extension that supports fast hypothetical deletions and scaling updates for Select-Project-Join-Aggregate-Union (SPJAU) queries, achieving low latency and high throughput (>1M hypothetical deletions or scaling updates per second)—orders of magnitude higher throughput than any prior approach. To showcase the expressiveness of FaDE’s API, we explore two data-driven applications. First, we use FaDE’s whatif() API to update linked visualizations by modeling updates as hypothetical deletions, effectively removing the influence of non-selected tuples. Second, we use FaDE to build an interactive explanation engine that uses the whatif() API to apply a set of hypothetical updates (12.7K interventions in <2ms) to find the best predicates that explain a set of outliers with just a few lines of code. Finally, explanations vary in the base query, metric, and space of candidate predicates. To demonstrate FaDE’s versatility, participants can use FaDE as a DuckDB extension through a Python Notebook, reuse workload from previous applications, and explore FaDE’s whatif() API.
