Rebalancer is a resource-allocation library developed at Meta for assigning objects to containers under constraints and optimization goals. Its C++ core runs in one process with multiple threads; applications can use C++ or Python interfaces.[2] Meta announced its open-source release on September 21, 2026, after more than nine years of internal use.[1]

Modeling an assignment

The model describes what can be placed, where it can go, and the properties relevant to placement. The same task-to-server problem can express limits on individual servers and requirements across racks.[3]

Construct	Meaning	Task-placement example
Object	Item being assigned	Task
Container, also called a bin	Assignment destination	Server
Dimension	Numeric object or container property	Task CPU requirement or server CPU capacity
Scope	Grouping of containers	Servers belonging to a rack
Partition	Grouping of objects	Tasks belonging to jobs

Dimensions can depend on both the object and its destination. Object groups need not be disjoint.[3]

Goals and constraints

Reusable specifications, or specs, describe placement requirements and quantities to optimize. The reference catalog contains more than 25 specs. Some support both goals and constraints; others support only one.[4]

Spec	Example purpose	Supported role
Capacity	Keep utilization within limits	Both
GroupCount	Bound a group's utilization within each scope item	Both
Balance	Equalize utilization across scope items	Goal
AvoidAssignments	Exclude particular placements	Constraint
ToFree	Drain listed containers	Both
DisasterRecoveryCapacity	Reserve capacity for correlated failures	Both

A constraint's policy matters when the initial assignment is invalid.[5]

Policy	Treatment
DEFAULT	Preserve initially satisfied constraints. Repair existing violations as high-priority goals while prohibiting deterioration beyond their initial level.
HARD	Require satisfaction; inability to satisfy the constraint is an error.
SOFT	Penalize violations, which can remain or worsen when the overall objective improves.

Multiple goals can share a weighted sum or occupy ordered priority buckets. Buckets are compared lexicographically: a lower-priority goal can improve only when earlier buckets tie. A large weight alone does not establish strict priority.[6]

Expression graph and solvers

Rebalancer translates specifications into a directed acyclic expression graph. The OSDI 2024 paper describes a representation that grows with objects plus bins, rather than the object-bin product of a direct mixed-integer formulation. The graph separates problem representation from the choice of solving algorithm.[7]

Solver family	Approach	Important limitation
Local search	Improve an initial assignment through incremental moves	Can stop at a local optimum rather than the best possible assignment
Optimal solver	Translate the model for FICO Xpress, Gurobi, or open-source HiGHS	Large problems can exhaust resources; a time-limited result need not be optimal

The model can remain unchanged when switching between these families.[2][8]

Local search is the default. Its configurable move types determine which neighboring assignments it evaluates. Time and applied-move limits provide stopping conditions; without configured stopping limits, it searches until reaching a local optimum. Discontinuous requirements can make incremental search get stuck.[9]

Production use at Meta

Meta uses Rebalancer for database-shard placement, allocating servers to services, traffic routing, serverless-function locality, and balancing online machine-learning training across regions. Its September 2026 release post reported the following operating scale.[1]

Reported measure	Value and scope
Daily assignment problems	Roughly 40 million
Distinct problem formulations	More than 30
99th-percentile solve time	12 seconds for a problem with 265,000 objects and 3,200 bins
Large-problem average	171 seconds for problems exceeding 1 million objects with 5,000 bins; more than 3,400 such runs

These figures describe Meta's workloads, not a guaranteed runtime for arbitrary models.[1]

Research and evaluation

Neeraj Kumar and eight coauthors presented Rebalancer at OSDI 2024. In one Azure-data placement experiment involving 10,000 pods and 500 nodes, the paper reported 94.6% placement with an optimal MIP solver, 92.8% with SINGLE_GREEDY, and 93.2% with SINGLE_RANDOM. SINGLE_RANDOM ran faster but placed fewer pods than the MIP baseline in that experiment.[7]

The paper's separate DCM comparison used published DCM measurements rather than running both systems on identical hardware, because of security and dependency restrictions.[7]

Explorer and distribution

Rebalancer Explorer is a web interface for inspecting solver inputs, assignments, goals, constraints, and solution scores. Its Next.js frontend communicates through a JSON proxy with a C++ Thrift backend. Docker Compose can start these components with example problem bundles. The proxy URL and bearer token are configurable; the token is required unless proxy authentication has explicitly been disabled.[10]

Rebalancer is distributed under Apache 2.0. Python installation uses pip install rebalancer; the repository also provides platform-specific native builds and packages. Building the C++ library from source does not itself create the Python package.[2]

See also

Meta AI

References

Engineering at Meta. Open-Sourcing Rebalancer: A Generic, High-Performance Library for Solving Assignment Problems. September 21, 2026.
Meta. Rebalancer repository and README. GitHub.
Meta. Core Concepts. Rebalancer documentation.
Meta. Goals and Constraints. Rebalancer documentation.
Meta. Constraint policy. Rebalancer documentation.
Meta. Goal priorities. Rebalancer documentation.
Kumar, Neeraj, et al. Optimizing Resource Allocation in Hyperscale Datacenters: Scalability, Usability, and Experiences. OSDI 2024, pp. 507-528.
Meta. Solver Overview. Rebalancer documentation.
Meta. Local Search Solver. Rebalancer documentation.
Meta. Rebalancer Explorer. Rebalancer documentation.

Editing: Rebalancer