Alby96andClaude Opus 5
91f9ac0fbf
Completes milestone 3: frames render over the map, play as a loop, and can be scrubbed, with the legend drawn from the manifest rather than a constant that could drift from what the worker actually rendered. The overlay alternates two MapLibre image sources. Updating one source in place flickers, because the layer briefly shows a half-written texture; adding every frame as its own layer avoids that but pins them all in GPU memory, and twenty 512x512 RGBA frames is about 80 MB. Two buffers cost the same whether the timeline holds six frames or sixty. Platform-channel work is serialised because `show` is called faster than the round trip completes during playback, and overlapping updates would swap visibility out of order and strobe. Prefetching loads a window around the playhead, nearest first and forward before backward, since playback moves forward and that frame is needed soonest. `FrameCache` is byte-budgeted rather than entry-counted because frame size tracks how much precipitation is on screen, and it evicts by distance from the playhead: plain LRU would keep frames the prefetcher touched a moment ago even after the playhead moved to the far end of the timeline. Also adds DpcRadarSource, which reads published frames from our CDN and never from the DPC API. Without it, the `dpc` adapter would have had to fall back to mock, putting demo frames on screen under the label of live data — exactly the confusion the adapter split exists to prevent. It now fails naming the missing setting instead. Running it on the emulator caught three things the tests had not: - The notifier wrote to `state` from inside `build()`, which Riverpod rejects as an uninitialised provider. That broke startup, not just tests. - Eight-month-old demo frames rendered as "Aggiornato 342535 minuti fa". The age formatter now steps up to hours and days. - Demo mode sat permanently behind a stale-data warning and so never showed the working state it exists to demonstrate. MockRadarSource now shifts the bundled timestamps onto the present, leaving images, order and spacing untouched, so the timeline behaves exactly as it would on live data. Corrects docs/stack-decisions.md, which described a disk cache that was not built: mock frames already live in the asset bundle, so a disk layer belongs with the network adapter where it would save a real request. Verified: analyze clean, 129 tests passing, and on the emulator the loop advances, wraps, and reports "Aggiornato ora". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
5.6 KiB
Stack decisions
Each entry records what was chosen, what it was chosen over, and why. Revisit an entry only with a reason that invalidates its rationale.
Flutter, single codebase
Android ships first, iOS follows on the same code. Platform-specific work stays behind
interfaces (core/platform/) so the iOS port is additive rather than a rewrite.
Toolchain pinned during setup: Flutter 3.47.3 stable / Dart 3.13.3,
compileSdk/targetSdk 36, minSdk 21 (the floor imposed by maplibre_gl),
JDK 21.
State management — Riverpod without code generation
flutter_riverpod with hand-written Notifier / AsyncNotifier classes.
Chosen over: Bloc (more ceremony than this app's state needs), plain setState
(the radar timeline, the frame cache and the consent flow all share state across
screens), Riverpod with riverpod_generator.
Dropping code generation keeps build_runner out of CI and out of every edit-run
cycle. The generator's benefit — less boilerplate on providers — is small at this
size, and its cost is paid on every build. Reconsider if provider count passes ~40.
Networking — Dio
dio with interceptors for: the mandatory MET Norway User-Agent, retry with
exponential backoff, and timeouts.
Chosen over http, which has no interceptor model — the User-Agent obligation is a
licensing requirement, so it belongs in one enforced place rather than at each call
site where it can be forgotten.
The DPC origin header lives only in the Python worker. The app never talks to
DPC directly.
Models — hand-written fromJson
Chosen over freezed + json_serializable. The model set is small (radar manifest and
frames, forecast series, CAP alerts, region config) and the parsers are covered by
tests against real fixtures. Same rationale as Riverpod: no build_runner.
The tests, not the generator, are what guarantee the parsing is right — fixtures captured from the real endpoints catch schema drift that codegen would not.
Map — MapLibre GL
maplibre_gl 0.27.0 (Flutter 3.29+, Android API 21+, iOS 13+).
Chosen over flutter_map, which renders tiles in Dart. Radar animation redraws a
full-viewport image several times a second; a GPU-composited native renderer holds
frame rate where a Dart canvas does not.
The base map style URL comes from MAP_STYLE_URL. With no key configured the app
loads a minimal local style — flat background plus the Piedmont boundary from a
bundled GeoJSON — so development, tests and the demo mode all work offline.
Radar frame rendering — image source, double buffered
The worker publishes each frame already cropped to the region bounding box and
reprojected to EPSG:3857, so the four corners of a MapLibre LatLngQuad are exact and
no client-side warping is needed.
Animation uses two image sources, A and B: while one is visible the next frame is decoded into the other, then visibility swaps.
Chosen over: adding all ~20 frames as layers with opacity 0 (constant GPU memory matters more than the saved swap — 20 frames at 1024×1024 RGBA is ~80 MB resident), and over updating a single source in place (visible flicker during decode).
Frame cache — custom in-memory LRU
FrameCache: an in-memory LRU of PNG bytes with a byte budget, plus retainOnly so
the timeline can drop everything outside the window around the playhead.
The budget is in bytes rather than entries because frame size tracks how much precipitation is on screen — a clear sky compresses to almost nothing, a storm does not — so an entry count would let a stormy loop use several times the memory of a calm one.
Chosen over flutter_cache_manager, which does not expose the eviction control the
scrubber needs. Plain LRU keeps frames the prefetcher touched a moment ago even after
the playhead has moved to the far end of the timeline, so eviction is driven by
distance from the playhead, not by access time.
Deliberately not persistent. Mock frames already live in the asset bundle, and network frames are re-fetched from the CDN. A disk layer belongs with the network adapter, where it would actually save a request, and is worth adding once real CDN frames are flowing.
Region configuration
Everything region-specific lives in app/assets/regions/<id>.json: bounding box, map
centre and zoom limits, alert zones, active data sources, attribution strings. Adding
a region is a new JSON file plus its assets — no Dart changes.
Piedmont bounding box, padded: [6.55, 43.95, 9.30, 46.55] (W, S, E, N).
To be refined against the ISTAT geometry once that dataset is confirmed.
Backend — Python worker on a VPS
GDAL and HDF5 cannot run on Supabase Edge Functions (Deno), so the worker runs as a scheduled process on a VPS or Cloud Run.
rasterio 1.5.1 publishes Windows wheels for Python 3.14, so the worker also runs
natively on the development machine — no Docker or WSL needed to iterate.
Trigger: the DPC WebSocket push channel wss://radar-wss.protezionecivile.it, with a
5-minute cron as fallback. Polling DPC on a timer is the fallback, never the norm.
Output goes through a Publisher interface: LocalPublisher writes to backend/out/
for development, S3Publisher targets R2 or Supabase Storage once credentials exist.
Rain notifications — client-side topic subscription
The worker computes rain per geographic cell and publishes per-cell state. The app subscribes to FCM topics named after cells, from the device.
The server therefore never learns any user's position — not precisely, not even by cell. There is no user table to leak, and Data safety can honestly declare that no location is transmitted or stored.