Choosing a town moved the camera and then left the reader to work out which of the settlements now on screen was the one they asked for. A town has no precise position, so the map now marks its centre: a white disc under a coloured dot, drawn as a style layer so the native renderer keeps it pinned through every pan and zoom. The radar frames are inserted below it, because a band of rain must not paint over the one thing that says which town this is. A searched municipality becomes a PointTarget rather than being thrown away as "the whole region". It is marked, the recentre button returns to it, and the chip names it - but it is never persisted, so looking something up no longer costs the user the place their app opens on. That was a real defect: searching went through select(FreeTarget), which wrote null over the stored preference. Following the device was broken outright. MapLibre reports a camera animation the app started exactly as it reports the user grabbing the map, so engaging follow and then animating to the position cancelled the follow it had just started; the resulting FreeTarget then re-framed the whole region. Follow now moves the camera first and switches tracking on second, and FreeTarget no longer moves the camera as a reaction to the state changing - framing the region is an action, so panning away while following keeps the view the user panned to. Verified on the emulator with a fix in Turin: a searched town is marked and saveable, the app reopens on it with the marker in place, following centres on the blue dot, and panning stops the chase without yanking the map away. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
8.4 KiB
Architecture
Shape of the system
DPC radar API ──┐
│ (worker only: origin header, presigned S3, 5-min cadence)
ARPA CAP feed ──┤
▼
Python worker ──► object storage / CDN
(crop, reproject, manifest.json
colourise, render) frames/*.png
alerts.json
cells.json
│
▼
Flutter app ──► OpenFreeMap (base map tiles)
The app never talks to DPC or ARPA directly. Both would be rate-limited by thousands of clients, DPC presigned URLs expire in minutes, and the rasters are whole-Italy GeoTIFFs that a phone should not decode. The worker is the only client of those services, and it fans out through a CDN.
The base map is the one exception, and it is not our data: OpenFreeMap serves public OpenStreetMap vector tiles with no key and no limits, and proxying them through our own infrastructure would add cost and latency for nothing.
Monorepo
Nuvolari/
├─ app/ Flutter application (Dart package "nuvolari")
├─ backend/ Python worker
├─ docs/ this documentation
└─ tool/ verification and asset-generation scripts
App layers
lib/
├─ core/ cross-cutting, no feature knowledge
│ ├─ config/ Env (dart-define), feature flags
│ ├─ region/ RegionConfig + asset loader
│ ├─ net/ Dio client, User-Agent and retry interceptors
│ ├─ cache/ FrameCache (disk + memory LRU)
│ └─ l10n/ localisation plumbing
├─ data/ one folder per domain, each exposing an interface
│ ├─ radar/ RadarSource + Dpc/Arpa/Mock implementations + models
│ └─ alerts/ AlertSource + ArpaCap implementation
├─ features/ one folder per screen or coherent UI area
│ ├─ map/ timeline/ places/ settings/ alerts/ sources/
└─ l10n/ app_it.arb (template)
Dependencies point inwards: features depends on data, data depends on core,
core depends on nothing in the app. A feature never imports another feature.
What the map draws, and in what order
Three things sit on the base map, and the order matters:
place marker (GeoJSON source + two circle layers) <- always on top
radar frames (two image layers, double buffered)
base map (OpenFreeMap vector tiles)
PlaceMarker creates its layers when the style finishes loading, and RadarOverlay
inserts its frames below them with addImageLayerBelow. Order is not left to whichever
attaches first: the radar attaches when the first frame arrives, which can be before or
after the user picks a place, and a band of rain painted over the marker would hide the
one thing on screen that says which town was chosen.
The marker is a style layer rather than a widget in the Stack over the map. The native
renderer keeps a layer pinned to its coordinates through every pan and zoom; a widget
would need repositioning from Dart on each camera frame, one asynchronous coordinate
conversion at a time, and would swim behind the map while it moved.
Following the device
Two behaviours of MapLibre shape how FollowUser works:
- A camera animation the app starts is reported as tracking dismissed, exactly like the user grabbing the map. So follow is engaged by moving the camera first and switching tracking on second. Done the other way round, the animation that brings the camera to the user cancels the following it was meant to start.
- Dropping out of following lands in
FreeTarget, which is also the state the user reaches by choosing "the whole region". SoFreeTargetmust never move the camera as a reaction to the state changing — that would throw away the pan the user just made. Framing the region is an action: the opening view, the menu entry, the recentre button.
The adapter seam
abstract interface class RadarSource {
Future<RadarManifest> getLatestManifest();
Future<List<RadarFrame>> getFrames();
}
Three implementations:
| Implementation | Status | Purpose |
|---|---|---|
MockRadarSource |
active | Synthetic frames from assets. Runs with no network and no credentials — the default in tests and in demo mode. |
DpcRadarSource |
active | Reads manifest.json and PNG frames from our CDN. |
ArpaRadarSource |
disabled stub | Placeholder until ARPA authorization exists. Throws if constructed while its feature flag is off. |
The active source is resolved from the region config plus a runtime flag, so switching
sources is configuration, never a code change. AlertSource follows the same pattern.
Because MockRadarSource is a first-class implementation rather than test scaffolding,
the whole UI — timeline, scrubbing, prefetch, cache eviction, degraded states — is
exercisable offline.
Data contract
manifest.json, published by the worker and consumed by the app:
{
"region": "piemonte",
"product": "VMI",
"generatedAt": 1758706260000,
"bbox": [6.55, 43.95, 9.30, 46.55],
"crs": "EPSG:3857",
"frames": [
{ "ts": 1758706200000, "url": "frames/VMI/1758706200000.png" }
],
"legend": {
"unit": "dBZ",
"stops": [{ "value": 5, "color": "#4FA3D1" }]
},
"attribution": "Radar-DPC — CC BY-SA"
}
Frame URLs are relative to the manifest so the whole tree can be moved between hosts. The legend travels with the data: the app draws whatever the worker produced rather than hardcoding a palette that could drift from the rendering.
Degradation
Failure is normal here — the worker can be behind, a frame can be missing, the phone can be offline. The rules:
- Manifest unreachable → keep the last good manifest from cache, show a "dati non disponibili" banner with the age of the newest frame.
- Individual frame missing → hold the previous frame in the timeline; never a blank map.
- No frames at all → the base map and the alerts still work; only the radar layer is empty.
- Base map tiles unreachable → MapLibre draws what it has; the radar overlay is positioned geographically, not relative to the tiles, so it stays correct.
- Animation stops when the app leaves the foreground (
AppLifecycleState) so a backgrounded app never burns battery prefetching.
The banner always states when the data is from. Stale radar shown as if current is worse than no radar.
Privacy by construction
No user location ever reaches a server. Rain notifications work by the device subscribing to FCM topics named after geographic cells — the subscription happens on the device, so the backend holds no user records at all. With no advertising and no forecast provider, nothing about the user leaves the device except the area the map is looking at, which is inherent to any hosted base map.
Launcher icon
The icon is a white cloud on the app's seed blue #1F6FB2, so the launcher and the
interface are recognisably the same product. It is drawn, not hand-edited:
tool/generate_icon.py renders it from signed distance fields — a union of three
circles and one rounded box — and writes two 1024px sources into app/assets/icon/.
Those sources are not listed in the pubspec assets: block: they feed the build and are
never bundled into the app.
flutter_launcher_icons fans them out to the Android densities, the adaptive icon, the
Android 13 monochrome layer and the iOS set, all of which are committed.
To change it:
python tool/generate_icon.py
cd app && dart run flutter_launcher_icons
Two things to know before doing that.
The generator owns the framing. An adaptive layer is 108dp of which only the central
72dp survives the launcher's mask, so ADAPTIVE_WIDTH is measured against that safe zone
and adaptive_icon_foreground_inset is set to 0. Left at its default of 16 the tool
would inset the layer a second time and the icon would read visibly smaller than every
other one on the home screen.
flutter_launcher_icons 0.14.4 corrupts an unrelated Xcode setting. ios.dart:340
rewrites the value of any line containing ASSETCATALOG, so it writes AppIcon into
ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS, which is a boolean.
Check git diff app/ios/Runner.xcodeproj/project.pbxproj after running it and restore
those lines to YES. It also minifies AppIcon.appiconset/Contents.json onto one line;
re-indenting it keeps the file reviewable.