About DeciDB

A research collaboration between the HUDA Lab at NYU Abu Dhabi and UMass Amherst, bringing optimization directly into SQL.

Research Foundation

DeciDB builds on a decade of research into in-database constrained optimization:

DeciDB: In-Database Processing of Decision Queries

Anh L. Mai, Filip Milisov, Hatim Rehmanjee, Azza Abouzied, Peter J. Haas, Alexandra Meliou

Submitted to CIDR 2027

Read Paper

Scaling Package Queries to a Billion Tuples

Anh L. Mai, Pengyu Wang, Azza Abouzied, Matteo Brucato, Peter J. Haas, Alexandra Meliou

Accepted to VLDB 2024

Read Paper

Scalable Computation of High-Order Optimization Queries

Matteo Brucato, Azza Abouzied, Alexandra Meliou

Accepted to ACM, 2019

Read Paper

Scalable Package Queries in Relational Database Systems

Matteo Brucato, Azza Abouzied, Alexandra Meliou

Accepted to VLDB 2016

Read Paper

The Team

NYU Abu Dhabi

Prof. Azza Abouzied

Professor of Computer Science

Anh Mai

PhD Student

Filip Milisov

Research Scientist

Hatim Rehmanjee

Research Scientist

UMass Amherst

Prof. Alexandra Meliou

Professor of Information and Computer Sciences

Prof. Peter J. Haas

Professor of Information and Computer Sciences

HUDA Lab

Human Data Interaction Lab

The HUDA Lab focuses on making data systems more accessible, interactive, and intelligent.

Research Areas:

  • • Query interfaces for non-experts
  • • Visual data exploration
  • • Automated data analysis
  • • Optimization in databases
Visit HUDA Lab Website

Technical Details

Built With

  • •
    DuckDB — Embeddable analytical database engine
  • •
    Gurobi — Primary MILP/QP/QCQP solver, auto-detected at runtime (commercial; recommended for performance)
  • •
    HiGHS — Bundled open-source LP/MIP and continuous convex QP solver; handles supported models when Gurobi is unavailable

Current Status

Distribution: Tagged releases provide command-line binaries and Python packages
Branch: master (active development)
Test Coverage: 1,600+ Python cases across 84 test files, 9 focused C++ test files, and a 94-query golden corpus
Target Scale: 1M rows in the standard TPC-H benchmark (exact solving with Gurobi)

Future Development

  • •
    Sketch-Refine Algorithm

    Scale beyond exact solving to very large datasets with bounded approximation error

  • •
    Broader Non-Linear Coverage

    Extend automatic linearization beyond the current ABS, MIN/MAX, <>, IN, and McCormick bilinear forms; full QCQP/SOCP on HiGHS