Is this using curl or sqlite3 directly? Is it getting the whole file, or is it "cloud optimized"?

We use our NetworkFile implementation that uses curl underneath and range request, and plug it into sqlite3 I/O layer. That way we can have piece wise access to the parts of the file we need, and benefit from the local caching in our ~/.local/share/proj/cache.db of the chunks we have already read.

From my test on a synthetic grid of 500x500 nodes, with each square divided into 2 triangles, that generates a GeoPackage file of ~ 64 MB. The first access to extract a triangle involves ~ 20 network requests of 16 KB each. Accesses to other parts (not close to each other) of the file afterwards take between 6 and 14 requests.

It seems that FlatGeoBuf spatial index (packed Hilbert RTree) would involve less requests, but FlatGeoBuf supports only one layer per file, and we need 2 layers. The FlatGeoBuf files would also be larger than the GeoPackage one (presumably because both the triangle geometry and the RTree need to be stored, whereas with GeoPackage we only store the RTree of the bbox of the triangles, but not the geometry of the triangles themselves, just the indices to the vertices), although file size is probably less of a concern. We could possibly have a .tar of 2 flatgeobuf files? I'm not sure if that's worth the complication.