FB-11 — Route Planner map still shows no tiles

Depends on — · Size M/L · Status Not started

Goal

Opening the Route Planner (the "+" button, or tapping any existing route) must show a real, detailed map underneath the pins. Today it shows a flat, featureless area.

Context

Direct user feedback (docs/FEEDBACK.md):

Tapping "+" to create a new route still shows a blank map with no tiles/detail rendered -- you cannot see streets or anything to tell you where pins are being dropped. A previous attempt at this fix (FB-07) did not actually resolve it on-device.

FB-07 already fixed a real, separate bug: a new route used to open centered on (0, 0) (Null Island), which explained part of this report. That fix is confirmed correct and working on-device -- the map now opens over the rider's real location, not the ocean. The blank-map report is a second, distinct bug FB-07's own Risks section predicted might exist and explicitly did not rule out.

An investigation this round ruled out the two most likely-looking causes:

Not a second Riverpod container. lib/main.dart has exactly one ProviderScope for the whole app. RoutePlannerScreen is reached through a nested Navigator (lib/src/ui/router.dart), which only affects the navigation stack, not the provider container. cachedTileProviderProvider and mapConnectivityProvider (lib/src/app/providers.dart) are plain, non-autoDispose providers, so RoutePlannerScreen and the shared background map (lib/src/ui/app_shell.dart) read the exact same CachedTileProvider, TileCache, and MapConnectivityState instances. There is no isolation between them.
Not stuck skeleton mode. MapConnectivityState (lib/src/tiles/map_connectivity.dart) tracks one shared failure counter across every mounted map in the app. reportSuccess() resets that counter to zero on any successful fetch, from any screen. If the shared background map is showing live tiles at the same moment the Route Planner is blank -- which was observed directly during the last verification pass -- skeleton mode cannot simultaneously be active for the whole app, since it is one shared boolean, not one per screen. Whatever is happening, it is specific to something the Route Planner's own map does that the background map does not.

One real, evidenced hazard was found in lib/src/tiles/tile_cache.dart, put():

@override
Future<void> put(TileKey key, Uint8List bytes) async {
  await _ensureLoaded();
  final name = _fileName(key);
  _manifest.remove(name);
  if (_totalBytes() + bytes.length > maxBytes) {
    await _evictUntilFits(bytes.length);
  }
  await _tileFile(key).writeAsBytes(bytes);
  _manifest[name] = _Entry(bytes: bytes.length, lastAccess: _clock++);
  await _saveManifest();
}

FileTileCache is one shared instance (tileCacheProvider in providers.dart), used by every TileLayer in the app. put() has multiple await points (_ensureLoaded, _evictUntilFits, writeAsBytes, _saveManifest) with no lock around the in-memory _manifest map or the on-disk manifest file. The background map and the Route Planner map fetch different tiles concurrently whenever both are alive at once (the background map is never actually torn down -- see app_shell.dart's comment on why it is one persistent instance). Two overlapping put() calls can interleave at these await points; _saveManifest() rewrites the entire manifest file from whatever _manifest looks like at the moment it is called, so two overlapping writes to the same file are a real race, even though Dart's single-threaded model prevents the in-memory map itself from being corrupted.

lib/src/tiles/cached_tile_provider.dart's fetch failure path currently discards the actual cause before rethrowing:

Future<Uint8List> _fetchAndStore() async {
  try {
    final response = await client.get(Uri.parse(url), headers: headers);
    if (response.statusCode != 200) {
      throw Exception('Tile fetch failed: ${response.statusCode} for $url');
    }
    ...
  } catch (_) {
    connectivity?.reportFailure();
    rethrow;
  }
}

The catch (_) block never records the URL, status code, or exception anywhere a developer could see it -- so there is currently no way to tell, from a real device, whether a Route Planner tile fetch is failing outright (and why), succeeding but failing to render, or something else entirely. test/route_planner_screen_test.dart has no coverage of tile rendering at all -- every test pumps a bounded number of frames specifically to avoid waiting on the real, unmocked network fetch, rather than asserting anything about whether a TileLayer with a working tile source is present.

Design

Two independent changes, both worth making regardless of which one turns out to be the actual fix, because both are real defects in their own right:

Make FileTileCache safe under concurrent use. Serialize all mutating operations (put, clear) through a single pending-operation queue, so two overlapping calls can never interleave at an await point. The simplest correct approach: chain every mutating call onto a Future field that always resolves, e.g.:
```
Future<void> _queue = Future.value();

Future<T> _serialized<T>(Future<T> Function() op) {
  final result = _queue.then((_) => op());
  _queue = result.then((_) {}, onError: (_) {});
  return result;
}
```
Wrap the bodies of put() and clear() in _serialized(...). Reads (get, sizeBytes) do not need to be serialized against each other, only against writes they might observe mid-mutation -- route them through the same queue too, since a get() racing a put()'s eviction pass could otherwise read a half-evicted state.
Log real tile-fetch failures. In cached_tile_provider.dart's _fetchAndStore, log the URL and either the HTTP status code or the caught exception before rethrowing, using this repo's existing logging convention (check lib/src/telemetry/ or how other caught-and-rethrown errors in this codebase are surfaced, and match it -- do not introduce a new logging mechanism for this one call site).

Implementation

Add the serialization queue to FileTileCache in lib/src/tiles/tile_cache.dart. Wrap put() and clear() bodies in it. Route get() and sizeBytes() through it too.
Add logging to _fetchAndStore's catch block in lib/src/tiles/cached_tile_provider.dart, matching this repo's existing logging pattern.
Add a concurrency test to test/tile_cache_test.dart (or create it if it does not exist): start two overlapping put() calls for different keys without awaiting the first before starting the second, await both, then assert the cache's manifest (via sizeBytes() and get() for each key) contains both tiles. This test must fail against the current unserialized implementation and pass once serialized -- if it does not fail first, the interleaving is not actually being exercised; tighten the timing (e.g. an artificial delay in a fake Directory/file layer) until it does.
Run the app on the Android emulator with the new logging in place. Set a mock GPS fix. Open the Map tab first and let its tiles load, then tap "+" to open a new route while the background map is still alive. Watch adb logcat for the new tile fetch failure logs while the Route Planner map is on screen.
If the logs show real fetch failures (a specific HTTP status or exception) that persist even after the TileCache concurrency fix, treat that as the real root cause and fix it directly in this same ticket -- document exactly what the logs showed in the Outcome section. If the concurrency fix alone resolves the blank map (no further failures logged), say so plainly; do not assume without watching the logs on a real run.
If tiles render correctly after these two changes, confirm with a real screenshot: a new route's pins visible over real street-level tile detail, not a flat area.

Acceptance criteria

Opening a new route via "+" shows real street-level tile detail under the pins, confirmed with a real on-device screenshot.
Opening an existing route with waypoints also shows real tile detail.
The new TileCache concurrency test fails without the serialization fix and passes with it.
The Outcome section states plainly what the on-device logs showed, and whether the concurrency fix alone resolved the blank map or a further root cause was found and fixed.
flutter analyze clean, flutter test green, test count only goes up.

Tests

test/tile_cache_test.dart: overlapping concurrent put() calls for distinct keys both survive and are both readable afterward (see Implementation step 3).
If a further root cause is found via the on-device logs (e.g. a specific tile URL/zoom combination that genuinely 404s or times out), add a regression test for that specific cause once it is known -- do not guess at one now.

Risks

The TileCache concurrency fix may not be the actual root cause -- the investigation that found it could not fully confirm it against a live repro. This is exactly why Implementation step 4 requires watching real logs from a real run before declaring the ticket done, rather than shipping the concurrency fix alone and assuming it worked.

Out of scope

FB-10's map-pan race, tracked separately. Any change to which tile provider or map style this app uses.

9.4 KiB Raw Blame History