Rethinking Last-Mile Routing hits 1M stops

The interesting claim here is not that VRP got “solved,” but that decomposition, caching, and boundary repair dominate once you leave toy-scale routing.

• –Constraint-aware clustering matters more than pure geometric partitioning because bad clusters create downstream infeasibility and boundary churn
• –Bounding route optimization per cluster keeps compute predictable, which is usually the difference between a workable pipeline and an unbounded search problem
• –The paper’s linear-ish scaling claim suggests that the expensive part of large VRP is often repeated orchestration work, not the routing heuristic itself
• –Reusing distance computations and reconciling cluster edges are the kind of unglamorous systems choices that decide whether large-scale routing is viable in production
• –The main caveat is that this is empirical scaling on a specific architecture and dataset, so the result is promising but not a universal complexity guarantee