Investigate NaN being produced internally during CRS work #2219

Activity

1. JaffaKetchup

   assigned

   mootw

   on Jun 30, 2026
2. JaffaKetchup

   added the

   Bug

   issue type on Jun 30, 2026
3. JaffaKetchup

   added

   P: 2 (soon™?)

   on Jun 30, 2026
4. aki1770-del commented on Jul 3, 2026

   

   aki1770-del

   on Jul 3, 2026

   More actions

   Hi @JaffaKetchup — took a look since this is squarely in an area I work in (a snow-driving nav stack on flutter_map).

   Root cause: getZoomScale(to, from) is crs.scale(to) / crs.scale(from). The default Epsg3857 extends CrsWithStaticTransformation, which doesn't override scale(), so it uses the base Crs.scale = 256 * math.pow(2, zoom). For normal zooms that's a finite number, so getZoomScale(z, z) is a clean 1.0 — the common case is fine. It only degenerates to NaN when scale(z) itself is 0, ±∞, or NaN (i.e. x/x with a degenerate x). Reproduced against the current code:

   scale(1017)  = Infinity   →  getZoomScale(1017, 1017)   = NaN   (∞/∞)
   scale(-1100) = 0.0        →  getZoomScale(-1100, -1100) = NaN   (0/0)
   scale(NaN)   = NaN        →  getZoomScale(NaN, NaN)     = NaN
   getZoomScale(14, 14)      = 1.0                                (normal — fine)
   

   On your suggestion (return 1 when the inputs are the same): it fixes the finite-extreme case and also catches ±infinity (since inf == inf is true) — but it misses the NaN case, because NaN == NaN is false in Dart, so the to == from guard is skipped and the NaN falls through. A complete guard would also short-circuit non-finite inputs (if (!toZoom.isFinite || !fromZoom.isFinite) return 1;).

   That said, the more interesting question is where a NaN/extreme zoom comes from — getZoomScale(z, z) can't be NaN for a normal z (it's always 1.0), so the NaN-during-interaction reports (e.g. #2199) suggest a NaN or runaway zoom is entering the camera upstream, probably from gesture/animation math, and getZoomScale is just where it first surfaces. A guard here stops the crash; tracing the upstream NaN would stop it at the source.

   Happy to put up a small PR with the guard + a regression test (I've got a local repro covering the identical-args finite/∞/NaN cases) if that'd be useful — or leave it to you. Whatever's easiest.
5. JaffaKetchup

   updated Priority Medium and Effort High on Jul 9, 2026
6. JaffaKetchup

   removed

   P: 2 (soon™?)

   on Jul 9, 2026