Outside My Window

Visualization app for French daily climatological data from Météo-France (data.gouv.fr).

Select a department, a weather station, and a year range to explore temperature, precipitation, and wind history, or compare the temperature distribution of two periods.

• Python 3.13+
• uv

data/stations.json is committed, so this is only needed when the station network changes:

uv run python scripts/build_station_index.py          # full: positions + record span
uv run python scripts/build_station_index.py --fast   # positions only, much quicker

The full run reads all three periods of every department (~1.5 GB of downloads, cached under data/cache/, one request at a time) because a station's first year is only knowable from the historical files. It records first_year / last_year / n_years, which the map uses to colour markers by record length. --fast reads the latest period alone and omits those fields; the map then falls back to a single colour. Either way the index lists the stations still reporting today.

Downloads use stable data.gouv.fr permalinks resolved from data/resources.json. Rebuild it only if the dataset re-issues resources with new ids:

uv run python scripts/build_resource_index.py

uv run python app.py

Then open http://127.0.0.1:8050 in your browser.

The landing page is a map of all stations. Click a station to view its temperature, precipitation, and wind history. Department data is fetched from Météo-France on first visit and cached locally under data/cache/.

Additional requirements: Node.js (includes npm)

npm install        # install Electron + electron-builder (one-time)
npm start          # launch the desktop window against this checkout

This starts a local waitress server on a free port and opens it in an Electron window. The Python sidecar is killed when you close the window.

Data is cached for 6 hours; press Ctrl+Shift+R (Cmd+Shift+R on macOS) to force-refresh the latest data immediately instead of waiting for the cache to expire.

The uv run python app.py workflow above still works for browser-based development.

npm run dist       # installer for the current platform, into release/
npm run dist:dir   # unpacked app only, into release/, for a quick check

npm run dist first freezes the Python server with PyInstaller into sidecar/omw-server/ (npm run build:server does that step alone), then has electron-builder package it as an NSIS installer on Windows, a dmg on macOS, or an AppImage on Linux.

The result needs neither Python nor uv on the target machine — the interpreter, the dependencies, the station index and the stylesheet are all inside the bundle. Expect ~600 MB unpacked, most of it Electron and Chromium. Downloaded weather data is written to the per-user application data directory, never next to the installed binary.

uv run pytest tests/ --cov=src --cov-report=term-missing

Browse every Météo-France station still reporting, across metropolitan France and the overseas departments. Markers are coloured by how long the station's record runs, so a 100-year series is visible at a glance. Search by name from the map or from any station page — the search is the one control that crosses department boundaries.

Every station page opens with what the station is doing now: the latest observation read against the 1991–2020 WMO standard normal for that same calendar day, the anomaly, and where the day ranks among every year of the record. Below it, the all-time records — hottest day, coldest night, wettest day, strongest gust, longest run of hot days.

Explore daily temperature (min/max band), precipitation, and wind for any station and year range. Switch between daily, weekly, and monthly granularity: temperatures average, precipitation totals (mm/week, mm/month), and gusts report the period's peak. Gaps in the record are drawn as gaps.

Track hot days (Tmax ≥ 30 °C by default, seven other definitions available), tropical nights (Tmin ≥ 20 °C) and frost days (Tmin < 0 °C) year by year. Years in which the station observed under 90 % of days are excluded and shaded grey rather than counted as cool ones. Optional trend lines show the long-term evolution with slope and R², computed on the fully observed years only.

Compare average Tmin and Tmax by month of year, either over the full record or broken down by decade — pick which decades to overlay rather than reading all of them at once.

Overlay the smoothed probability density of daily Tmax and Tmin for two year ranges over the same season window (e.g. 1 Jun – 31 Aug, 1961–1990 against 1995–2024), so the shift of the whole distribution — and of its hot tail — is directly readable. Curves are Gaussian kernel density estimates with a Silverman bandwidth. A stats line below gives the mean, median and p90 of each period plus the shift between them.