The hard part of million-stop routing was not the route optimizer

Rethinking Last-Mile Routing at Scale Near-linear planning on commodity hardware This repository contains a technical report describing a practical architecture for large-scale last-mile route optimization. The core idea: at million-stop scale, vehicle routing stops being only an optimization problem and becomes a systems problem involving partitioning, boundary repair, graph reuse, bounded route-level optimization, and orchestration. Paper 📄 Download the PDF Main results Under a shared external measurement protocol based on OSRM and Google Maps, using the public Amazon Last Mile Routing Research Challenge dataset: 23.3% less measured distance relative to the released Amazon baseline routes 11.1% fewer routes 17.59% mean depot-level distance reduction 1,000,000 stops processed in ~20 minutes on commodity hardware What this is This is a systems-oriented technical report about making very large last-mile routing workloads practical. The architecture combines: parallel constraint-aware clustering constraint-aware vehicle allocation distributed boundary rebalancing bounded route-level optimization localized graph and distance reuse What this is not This report does not claim to reproduce Amazon’s internal routing objective. The benchmark compares released Amazon route files and generated route files over the same stops using a shared external OSRM/Google measurement protocol. The goal is to compare routing structure under a reproducible external metric, not to infer or match Amazon’s proprietary operational objective. Validation A separate verifier repository is available here: https://github.com/vizzito/last-mile-route-verifier It is intended to help inspect and reproduce the external distance comparison protocol. Why this matters Many routing APIs and logistics systems work well at small or moderate request sizes, but require manual pre-partitioning, strict input caps, or large infrastructure for very large workloads. This report explores an alternative: organizing the routing problem into bounded, composable stages so that workloads from thousands to one million stops can be planned predictably on commodity hardware. Citation @misc{vizzolini2026rethinking, author = {Martin Vizzolini}, title = {Rethinking Last-Mile Routing at Scale: Near-Linear Planning on Commodity Hardware}, year = {2026}, howpublished = {Technical report}, url = {https://github.com/vizzito/last-mile-optimizer-paper/releases/tag/v1.0.0} }

This excerpt is published under fair use for community discussion. Read the full article at GitHub.