Shapefile vs GeoJSON for road network updates: what works in practice

If you've ever had a vendor hand you a road update and the first thing you had to do was figure out what FUNC_CLS or RTE_TYP actually meant, you already know why this question keeps coming up. Shapefile and GeoJSON are the two formats you'll see most often when a road diff lands in your inbox, and they behave differently enough that picking wrong costs you real QA time.
Shapefile: the format everyone's toolchain already speaks
Shapefile is still the default a lot of production pipelines expect, especially if your editing workflow runs through older desktop GIS or a legacy ingest script. It's well understood, every tool opens it without a plugin, and a lot of municipal and state DOT exchange formats still specify it by name.
The catches are the same ones that have annoyed GIS people since the format shipped in 1998. Field names cap at 10 characters, so FUNCTIONAL_CLASS gets truncated to something like FUNCTIONAL and you're left guessing whether that matches the schema your conflation script expects. It's not one file but a bundle (.shp, .shx, .dbf, and usually .prj), which means a dropped sidecar file breaks the whole delivery and nobody notices until the load job throws an error. The attribute table uses dBase encoding, so anything outside basic Latin, an accented road name in a French-Canadian dataset, a transliterated name in a border region, can come through mangled. And there's a 2GB file size ceiling per component that you will eventually hit if a diff covers a large metro at high segment density.
None of this makes shapefile unusable for a road diff. It makes it a format you need to validate carefully before you trust it in an automated merge.
GeoJSON: easier to diff, easier to break silently
GeoJSON solves most of shapefile's structural problems. It's one file, UTF-8 by default, and attribute names aren't truncated, so a nested property like {"road_class": "arterial", "oneway": true} comes through exactly as written. Because it's plain text, you can diff two deliveries with a standard version-control tool and see exactly which segments changed, which is a real advantage when you're doing a quarterly reconciliation and want to know whether a vendor's new file only touched the areas they claimed to touch.
The tradeoffs run the other way from shapefile's. There's no enforced schema, so nothing stops one delivery from using "lanes": 2 and the next from using "lanes": "2", and your ingest script needs to catch that itself. Large files parse slower than an indexed shapefile or geopackage, because GeoJSON has no spatial index built in. Winding order on polygons matters more than most editors realize, though for a road network you're dealing with line geometry, so that particular gotcha mostly doesn't apply to you.
Which one to ask for
For a routing graph specifically, where you're merging in new and changed segments against an existing network rather than replacing the whole thing, GeoJSON's field-name fidelity and human-readable diffs usually win. You can open the file, see exactly what attributes came with each segment, and spot a schema mismatch before it corrupts your graph. Shapefile still makes sense if your downstream tools are older and genuinely expect it, or if a partner agency's exchange standard already mandates it. Geopackage is worth a mention too since it fixes most of shapefile's limits while staying in a single file, but it's less universally supported by lightweight ingest scripts than either of the other two.
Whatever format you land on, a clean diff is correctly attributed, geometrically accurate, and scoped only to what changed since the last pass. Road Extraction delivers new and changed road segments from high-res imagery as a vector layer built for exactly this kind of merge, on a quarterly cycle, so your ingest script is checking a schema rather than guessing at one.
If quarterly road diffs are already part of your workflow, it's worth seeing what a clean one looks like.