Rebuilding a national logistics network in 14 weeks

Dispatch was one Rails monolith: 700 tables, no read replicas, and a routes table nobody would run a migration against without a maintenance window. Planning ran as an overnight batch, so a truck that broke down at 06:00 stayed wrong on every screen until the next morning — the floor covered the gap with a whiteboard and a phone. Peak season pushed p95 API latency to 2.4 seconds. On-call was absorbing sixty pages a month, and roughly half of them were the same deadlock on the same table.

We spent the first two weeks on the dispatch floor in Memphis rather than in the repo, which is where the actual problem showed up: the overnight batch was not just stale, it was distrusted, so the floor kept a second plan on paper. We split the monolith along three seams that already existed in the org chart — dispatch, telematics ingest, billing — and moved each behind a strangler proxy so traffic could shift a percent at a time and roll back in seconds. The new planner recomputes on every telematics event instead of once a night, reading a Kafka log the legacy system also writes to. That is the part that made the cutover boring: both stacks ran on identical input for nine weeks while we diffed their output every morning, until the diffs stopped being interesting.

We cut over in week 14, one region at a time, with no shipment held. p95 landed at 180ms, and a full network replan went from eight hours to 38 seconds — which changed how the floor works more than the latency did, because a replan became something you just try. Empty running fell from 10.3% of kilometres to 6.1% over the following two quarters, and paging is down to around four a month. We still run the platform under a 24/7 agreement. The whiteboard is gone.