Meta has open-sourced Rebalancer, the assignment-problem solver it says has been running across its infrastructure for more than nine years. The useful bit is not simply that another internal tool has escaped into public view: Rebalancer is designed to express and solve the messy allocation problems behind servers, traffic, workloads and other large-scale systems.
Meta AI Watch analysis
What happened
Meta says Rebalancer lets engineers describe objects, bins, constraints and objectives, then turns those descriptions into an expression graph that can be solved with either mathematical optimisation or local search. That means the same framework can model tasks assigned to servers, racks spread across fault domains, traffic routed between datacentres, or machine-learning workloads balanced across regions.
The project supports mixed-integer programming with commercial solvers such as Gurobi and FICO Xpress, as well as the open-source HiGHS solver. For much larger problems, its local-search approach explores possible moves directly rather than attempting to build an enormous global model. Meta says the local-search system can evaluate millions of candidate moves per second.
Meta reports that Rebalancer handles roughly 40 million assignment problems each day across more than 30 problem formulations. Its examples include shard placement, service allocation, traffic routing and serverless-function grouping, alongside less glamorous but relatable jobs such as assigning meetings to rooms and support tickets to engineers. The code is released under the Apache 2.0 licence, with documentation, a Python package and Rebalancer Explorer, a Dockerised web interface for debugging solutions.
Key findings
- Two solving modes
Local search targets very large problems, while a more traditional optimal solver suits smaller or moderate-sized cases. - Production scale
Meta says Rebalancer solves about 40 million assignment problems daily across its infrastructure. - Useful debugging tools
Rebalancer Explorer shows binding constraints, tests what happens when rules are relaxed and helps explain why an object landed in a particular bin. - Open-source reach
The Apache 2.0 release invites use in fields including healthcare, energy, transport, education and emergency response.
Why it matters
Assignment problems are everywhere, but they are rarely tidy. A system may need to spread services across failure domains, respect power and cooling limits, balance CPU and storage, and still avoid turning the resulting calculation into a small personal catastrophe for the operations team.
Rebalancer’s practical value is its attempt to separate the description of a problem from the machinery used to solve it. That could make sophisticated optimisation easier to reuse, inspect and adapt. The open release also gives researchers and engineers a concrete system to test rather than another promise that optimisation is happening somewhere behind a tasteful diagram.
Meta’s reported performance figures are company claims, not an independent benchmark. The framework can still produce very large mixed-integer models, and local search does not guarantee a globally optimal answer. Those are important limits, especially when an allocation decision affects reliability, cost or safety.
Our read
This is a worthwhile release because it exposes infrastructure logic that usually stays hidden behind giant services and even larger acronyms. The strongest idea is not raw speed but the modelling layer: reusable constraints and objectives could help teams express operational trade-offs without rebuilding a solver for every new problem.
Researchers should inspect the code and paper. Engineers should start with a bounded workload and compare the local-search result with a trusted baseline before handing over the keys to anything mission-critical. The first great open-source optimiser is still no substitute for knowing what you are optimising, a detail software has been quietly requesting since the dawn of spreadsheets.
What to watch
- Whether independent users reproduce Meta’s reported solve times and scaling results.
- How Rebalancer performs on problems outside Meta’s infrastructure environment.
- Whether the Apache-licensed code attracts contributions for new solving techniques and constraints.
- How often teams choose local search when a provably optimal solution is impractical. Meta’s engineering report is the evidence for the release, its capabilities and the usage figures. The specific WittyWires relevance is concrete: this is an open-source computing system that solves large-scale resource-allocation problems used in AI infrastructure and other production workloads.
Discussion spark: Should infrastructure teams prefer a fast, explainable local-search solution, or hold out for a provably optimal allocation even when the workload is too large for conventional solvers?
Sources and evidence
- Open-Sourcing Rebalancer: A Generic, High-Performance Library for Solving Assignment Problems (21 September 2026, 16:00 UTC)
not affiliated with, endorsed by, or operated by Meta