Abstract
Marketplace recommenders operate under two competing pressures: they must surface relevant candidates at low latency while respecting supply-side capacity, policy constraints, and ecosystem balance. A common failure mode of purely local candidate generation is overproduction of redundant or infeasible options that are later filtered by business rules or ranking-time caps, wasting computation and reducing coverage for the long tail of listings and sellers. This paper presents a scalable bipartite graph matching system that produces recommendation candidates as a constrained assignment problem between demand entities and supply entities. The core contribution is blocked parallel augmentation, a distributed augmentation mechanism that decomposes the graph into blocks, performs localized augmenting path discovery and tentative updates in parallel, and commits non-conflicting augmentations via deterministic arbitration. The design targets the operational characteristics of marketplaces: dynamic eligibility edges, high churn in supply, asymmetric degree distributions, and the need for controllable trade-offs between solution quality and throughput. The system integrates a capacity-aware matching objective with practical constraints including per-supply exposure budgets, per-demand candidate quotas, and soft penalties that discourage repeated exposure. The paper details the algorithmic structure, consistency model, partitioning strategy, and an implementation that supports both batch and incremental updates. Analytical considerations and trace-driven experiments demonstrate how block sizing, conflict rates, and degree skew influence convergence and resource usage, and how the approach improves feasible candidate yield and supply coverage under fixed compute budgets.