Meta AI Open-Sources Rebalancer: A C++ Assignment Solver That Runs About 40 Million Placement Problems a Day

Meta has opened up Rebalancer, a C++ library with Python bindings for assignment problems: figuring out what goes into which bin under a pile of constraints and goals. It ships under Apache 2.0, comes with docs, a PyPI package, and a debugging tool called Rebalancer Explorer. You can install it now with pip, though PyPI still labels the project Alpha.

The pitch here is not just “we built a solver.” Meta says the real problem was getting engineers to express messy policies in a form a machine can use, while also dealing with problems that are often NP-hard and too large for commercial tools. So Rebalancer splits the work in two: a specification layer for describing the problem, and a solver layer that actually tries to optimize it.

The spec side has three pieces: modeling constructs, an expression API, and a spec API full of predefined objectives and constraints. Meta’s own example treats tasks as objects, servers as bins, and racks as a scope, then layers on rules like CPU and storage caps, limits on job types per rack, and balance requirements across dimensions. The whole thing compiles into a directed acyclic expression graph.

From there, Rebalancer can go two ways. For smaller problems, it translates that graph into a mixed integer program for solvers like FICO Xpress, Gurobi, or HiGHS. For the big stuff, Meta uses local search directly on the graph, exploring moves object by object and pruning the search space. The company says that’s the path it uses for almost all large problems.

The scale numbers are the real headline. Meta says Rebalancer handles about 40 million assignment problems per day across more than 30 unique formulations. It reports a p99 solve time of 12 seconds on 265,000 objects and 3,200 bins, and says problems above 1 million objects and 5,000 bins average 171 seconds across more than 3,400 runs. Rebalancer Explorer exists because, apparently, watching a solver behave is easier than guessing why it made a weird choice.