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

Meta has open-sourced Rebalancer, a C++ library with a Python interface for assignment problems: deciding which objects go into which bins under a set of rules and goals. Inside Meta that means racks into datacenters, servers to services, tasks to servers and user traffic to regions. The company says the tool has handled resource allocation across Meta for more than nine years. It is released under Apache 2.0, installable from PyPI today, and ships with a debugging interface called Rebalancer Explorer.

The production numbers are serious. MarkTechPost, citing Meta, reports about 40 million assignment problems solved per day across more than 30 formulations, a 99th-percentile solve time of 12 seconds on problems with 265,000 objects and 3,200 bins, and an average of 171 seconds on problems above a million objects and 5,000 bins. This is the kind of infrastructure that keeps large AI and web services running without anyone noticing.

We are glad to see this released. Optimization is the unglamorous cousin of the AI boom, but every model served at scale depends on decisions about which machine runs what, and those decisions cost money and electricity when they are made badly. Meta's core design choice is sensible: separate how a problem is described from how it is solved. Engineers write constraints and objectives in a structured spec, and the same spec can be handed to a fast local search for huge problems or to a mixed integer programming solver for smaller ones. Meta says it often prototypes with the exact solver and moves to local search when the problem grows.

The part I keep coming back to is Rebalancer Explorer. Meta found that its modelers spent most of their time debugging solver behavior, so it built a tool that shows which constraints are binding and why a given object landed in a given bin. That is a small lesson with wide reach. Automated decisions are only trustworthy when someone can ask why and get an answer.

That matters more as the objects stop being servers. The release lists operational uses outside infrastructure: support tickets to engineers, meetings to rooms, desks to people. Once the bins hold human time, the objective function is a policy. Balance load and you may also be deciding who gets the worst shifts. Nothing about the tool is harmful on its own, but organizations that adopt it for people-facing assignments should treat the spec as something to review, not just tune.

Some caution for adopters. PyPI still labels the project Alpha, prebuilt wheels cover Linux x86-64 and macOS on Apple silicon, and the exact-solver path can lean on commercial engines such as FICO Xpress or Gurobi, although open-source HiGHS is supported. Mature alternatives exist too: MarkTechPost compares it to Google's OR-Tools, which covers more problem types, and Timefold, which targets scheduling on the JVM. Our read is that Rebalancer earns a look for anyone balancing large clusters, with the explainability tooling as the real differentiator.

The design was described in an OSDI 2024 paper, Optimizing Resource Allocation in Hyperscale Datacenters. Specs compile into a directed graph of utilization values, and constraints that the starting assignment already violates are treated as high-priority goals. Meta systems such as Shard Manager and its edge traffic balancer Taiji use this pattern.