Meta Open-Sources Rebalancer, the Placement Engine Behind 40 Million Daily Assignments

Meta has made public Rebalancer, an assignment solver that the company says has handled its large-scale placement work for more than nine years. The library answers a deceptively simple question: where should each piece of work go? That covers which data shard lives on which machine, which server hosts which workload, and how traffic is directed. By Meta's account, the system works through about 40 million of these assignment problems every day.

The library is written in C++ with Python bindings, and it offers two ways to solve a problem. The first is a built-in local search method that repeatedly improves an existing placement. The second passes the problem to a mixed-integer programming (MIP) solver, with Gurobi, FICO Xpress and HiGHS named as supported options. Teams that already license a commercial solver can keep using it, while HiGHS gives everyone else an open-source route with no license fee.

Placement problems show up throughout modern infrastructure. Whenever a distributed database rebalances its data or a scheduler moves containers, someone has to find a layout that respects capacity limits, failure-domain rules and the cost of moving things around. Many companies solve these with homegrown heuristics that rarely leave the building. Releasing a tool that has been tested at this scale under a permissive license gives smaller operators a concrete starting point, and it adds to Meta's recent pattern of publishing its infrastructure tooling.

Rebalancer is available now through pip under the Apache 2.0 license. It is a library rather than a finished product, so teams still need to translate their own constraints and objectives into the model. The announcement summary does not include independent benchmarks, so anyone weighing it against existing solvers should test it on their own workloads and read the project documentation closely.