Skip to content

Future work

Known issues and improvements deliberately deferred from the 2026-07-13 correctness sweep (see Data corrections).

Reliability

  • Connect timeouts are never retried. httpx2.ConnectTimeout is a subclass of httpx2.HTTPError but not of ConnectError/NetworkError/ProtocolError, so _get_json (plugins/_base.py) classifies it as ErrorCode.UNKNOWN, which is not in the client's retryable set — contradicting the documented retry policy, which promises that timeouts are retried. Fix: catch httpx2.TimeoutException first and map it to ErrorCode.TIMEOUT.
  • Plugin-raised httpx timeouts are labeled NETWORK. In client.py::_attempt_fetch the except httpx2.HTTPError branch catches TimeoutException before the generic handler that would classify it as TIMEOUT. Retryable either way; only logs/metrics are misattributed.

HTTP cache

  • CachingTransport._response_from_revalidation mutates the shared cache entry outside the lock; a concurrent reader between awaits can observe a half-updated entry.
  • CachingTransport with max_entries=0 would raise StopIteration on the first store. Unreachable via config (cache_max_entries is ge=1), but worth a guard if the transport is ever constructed directly.

Providers

  • OpenWeather hourly is_day is always None — hourly One Call entries carry no sunrise/sunset; it could be derived from the daily block.
  • Weatherbit imperial pressure — slp/pres are passed through unconverted for units="I"; verify against Weatherbit's docs whether imperial responses switch pressure to inHg.
  • Weatherbit sea-level pressure plausibility — one stored run shows slp up to 1074 hPa (implausible at sea level). Consider a provider-side plausibility filter or at least a data-quality flag in scripts/inspect_db.py.

Data model

  • Condition vocabulary gaps: provider texts like "Patchy rain possible" or "Thundery outbreaks" map through generic keywords; a per-provider code table (e.g. WeatherAPI's numeric condition codes) would be more precise than text matching.