Road/Extraction

Blog/ 2026-10-13

QA a road network vector diff before it touches production

A winding two-lane road with visible lane markings curving through green countryside under a clear sky

A vector diff is only useful if it merges clean. Hand a production team a batch of new and changed road segments with a dangling node or a mismatched one-way flag, and you haven't saved them time, you've handed them a debugging session with a deadline attached. The QA pass before merge is where that gets caught, and it's worth treating as its own step rather than something you do while merging.

What actually breaks a routing graph

Most merge failures trace back to a short list of repeat offenders:

  • Dangling nodes and stub segments. A new segment that doesn't snap to the existing graph at both ends leaves a node with nowhere to go. Routing engines don't fail loudly on this. They just quietly route around the gap, and nobody notices until a driver does.
  • Duplicate or overlapping geometry. Two segments covering the same centerline at slightly different offsets create a false fork in the graph. Depending on your routing engine, this can double-count distance or let a route jump lanes that don't connect.
  • Attribute mismatches at the seam. A changed segment that drops the one-way flag, loses a turn restriction, or arrives with a different road class than the segment it replaces will route fine in isolation and wrong in context.
  • Node-edge inconsistency. Segments that terminate mid-span instead of at a shared node, often from a simplification pass that wasn't graph-aware. Looks fine rendered, breaks the first time a routing engine tries to traverse it.
  • Z-level and crossing confusion. An overpass or underpass edit that doesn't carry the grade-separation attribute forward will flatten two roads into one intersection that was never supposed to exist.

None of these show up in a quick visual scan at normal zoom. They show up when something tries to route through them.

A merge-ready QA checklist

Before a diff goes anywhere near the production graph, run it through a staging pass:

  1. Geometry validation first. Check for self-intersections, zero-length segments, and duplicate geometry within the diff itself, before you even compare it to what's live.
  2. Snap and connectivity check against the existing graph. Every new segment endpoint should land on an existing node or a node introduced elsewhere in the same diff. Anything within a small tolerance but not snapped is a flag, not a pass.
  3. Attribute carry-forward audit. For changed segments, confirm road class, one-way direction, and turn restrictions either match the segment being replaced or have a documented reason for the change.
  4. Topology rebuild on a sandbox copy. Merge the diff into a staging copy of the graph, not production, and run your routing engine's own validation or a build pass over it. This is the step that catches what visual review can't.
  5. Spot-route test. Pick a handful of routes that cross the edited area and run them through before and after. If a route that used to take the arterial now detours three blocks for no reason, something in the diff broke a connection.
  6. Changelog cross-check. Confirm the count and location of changed segments in the diff matches what the delivery changelog says was touched. A mismatch here usually means something got clipped or dropped in transit.

Run this against a sandbox copy of the graph, never the live one. A checklist run in staging catches the break while it's cheap to fix, hours of work instead of a routing complaint and a rollback three weeks after the merge.

Where this fits in a quarterly update cycle

If you're running updates on a quarterly cadence, QA stops being a one-off fire drill and becomes a repeatable step with the same checklist every cycle. That's easier to staff and easier to hand off than reinventing the process each time a diff lands. Road Extraction's quarterly vector diff is built to drop into exactly this kind of staged merge: new and changed segments extracted from high-res imagery, delivered as a vector layer, with the QA and merge decisions left where they belong, with your production team.

Topology checks aren't glamorous work, but they're the difference between a graph that routes correctly and one that quietly doesn't, three weeks from now, in a corridor nobody's watching.

If a field survey backlog or a crowd-sourced edit queue is what's standing between you and current segments, it might be worth seeing what a quarterly diff could take off that list.

Start a project