Meta releases open-source assignment solver for efficient resource allocation

Meta has released Rebalancer, an open-source C++ library with a Python interface designed to solve complex assignment problems. The tool, available under Apache 2.0, handles over 9 years of resource allocation for Meta, including tasks like assigning servers to services and user traffic to datacenters. It supports deployment via pip, with prebuilt packages for Linux and macOS, though it is still classified as Alpha on PyPI. Rebalancer addresses challenges in usability and scalability by separating problem specification from solving. It compiles user-defined specifications into a directed acyclic graph, which can be solved using MIP solvers like Gurobi or HiGHS, or through local search methods. Meta uses MIP for smaller problems and local search for larger ones, often prototyping with MIP first. The tool also includes a debugging UI called Rebalancer Explorer to help modelers understand solver behavior. The project, detailed in an OSDI 2024 paper, competes with other optimization tools like OR-Tools and Timefold, offering a unique approach by supporting both local search and MIP. Meta highlights its ability to process up to 40 million placement problems daily, showcasing its efficiency in hyperscale environments.