Resiliency solvers: Optimizing workload placement with linear programming

As modern applications scale across massive global fleets, ensuring fault tolerance requires intentional planning and execution across infrastructure failure domains. Ensuring that sharded workloads are distributed effectively across power, cooling, and network boundaries is critical for maintaining high availability. However, manually balancing compute, memory, and storage constraints against these domains often leads to either wasted capacity or fragile placement.

This lightning talk explores a cost-effective approach to sharding resources using mixed-integer linear programming (MILP). By formulating workload placement as a constraint optimization problem and leveraging high-performance, licensed solvers like Gurobi and Xpress, infrastructure teams can automate complex placement decisions.

The session examines this mathematical framework using the eight queens and sudoku problems to show how infrastructure can meet application constraints. The framework applies to largest-scale services and helps make them resilient to failures without overspending on redundancy requirements.