papa-map

PapaMap — Wickeltische, die ein Vater erreicht.

A static map of places across 44 European countries — from Iceland to Cyprus, from Portugal to Ukraine — with a baby changing table, colored by whether a dad can actually reach it: green = accessible room (men's/unisex/dedicated/wheelchair), red = women's room only, grey = table exists but nobody has recorded which room — the call to action. Every grey pin deep-links to the same object on MapComplete so the missing answer becomes an OpenStreetMap contribution. OSM is the only data source and the only write destination; this repo owns no data and writes nothing to OSM itself.

Why: Google Maps, Apple Maps and Yelp have no changing-table attribute at all, and OSM's changing_table:location is the only open vocabulary on Earth recording which room a table is in — globally ~485 places record a women's-room-only table vs ~71 men's-room-only (~7:1). A full toilet-provision scorecard is impossible from open data (we checked), so this site measures what is measurable and turns the gaps into edits. The honest full story, including the exact classification rule: web/methods.html.

Setup

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

(Code targets Python 3.9+; from __future__ import annotations keeps the union type hints working on the system Python.)

Build the dataset

python -m pipeline.run   # Overpass + taginfo -> web/data/*.json + web/wickeltische/*.html

A fresh checkout sweeps all 16 Bundesländer plus Denmark and merges the results (an all-Germany area query dies at a ~60 s network idle cutoff, so Germany stays chunked per Land; Denmark is small enough to answer whole in one admin_level=2 query, ~15 s). ~5 min in total. papamap.de sweeps more than that — see the first bullet. Germany is no longer the only chunked country: France is swept as its 13 metropolitan régions for exactly the same reason, measured on 19 Aug 2026 as an empty reply at 60.14 s for the country whole.

• PAPAMAP_COUNTRIES picks the countries. The code default is de,dk — a fresh checkout and the test suite build Germany + Denmark, and that stays that way on purpose, so neither depends on a 44-country sweep. The deployment sets the variable instead: docker-compose.yml carries the full PAPAMAP_COUNTRIES list, so papamap.de sweeps 44 European countries — every European sovereign from Iceland to Cyprus except Vatican City (too small to survive the zero-objects check) and the transcontinental states. Codes are ISO 3166-1 alpha-2, so the UK is gb, not uk, and Kosovo is the user-assigned xk. All but two of them are one admin_level=2 area each, the way Denmark is. fr is 13 areas, one per metropolitan région, and de is the 16 Bundesländer — the only two chunked countries. So PAPAMAP_COUNTRIES=dk builds Denmark alone and de,dk,at,ch adds the German-speaking neighbours; an unknown code aborts rather than silently sweeping less. (The map's maxBounds is [[-32,27],[41,71.5]], each edge set by a real pin — the Azores west, the Canary Islands south, Ukraine east, Nordkapp north; web/app.js documents all four. Check it with any expansion, not after.)
• Every country since the 2026-08-18 ring is matched on name:en, because a country's own name can be several languages at once — Belgium is "België / Belgique / Belgien", Switzerland "Schweiz/Suisse/Svizzera/Svizra" — and a name= miss resolves to zero objects, which the build can only read as a failed sweep. Their region labels, on the leaderboard and in history.json, are therefore the English names. Germany and Denmark keep name=: Deutschland and Danmark are existing history keys. The United Kingdom keeps it too — its name and name:en are the same string. France's régions must use name=, where name:en would actively break: it is Bourgogne – Franche-Comté with an en dash, Ile-de-France with the accent dropped, and translated for four of the thirteen. The 13 are an allowlist, not a subdivision — the five overseas régions are admin_level=4 too, and sweeping them would put pins in the Caribbean, outside maxBounds where nothing can be panned to.
• A build of three or more countries names itself by count rather than by name: stats.json's area_key becomes countries_<n> (countries_44 for all 44), since the joined labels overflow the stats strip. One and two countries keep de / dk / de_dk.
• PAPAMAP_AREA_NAME + PAPAMAP_AREA_ADMIN_LEVEL select a single area instead (e.g. Hamburg / 4), and PAPAMAP_DISPLAY_AREA names the dataset in the stats strip.

Area pages

The same run writes one static page per swept area into web/wickeltische/ (git-ignored — they are build output). The map is a single URL for every country it sweeps, so a search for "Wickeltisch Bayern" — or "puslebord Danmark", or "table à langer Bretagne" — had nothing to match: the place names live inside a multi-megabyte GeoJSON that crawlers read as a download. These pages put each area's counts and its named places into HTML, and link back into the map at that area's extent via ?bbox=.

Each page is written in the language its readers search in, which is the whole point of their existing: German for the 16 Bundesländer (plus their index at web/wickeltische/index.html) and for Austria and Switzerland, Danish for danmark.html, English for united-kingdom.html, French for france.html — a hub over 13 per-région pages — and so on. The routing (including the inflected name forms prose needs: "in der Schweiz", "w Polsce") lives in config.COUNTRY_PAGES and pipeline/pages_l10n.py; slugs are the local names (belgie.html, cesko.html, oesterreich.html). The map's footer "Bundesländer" link is language-routed the same way (regionsHref in web/i18n.js): the Danish UI links danmark.html, the French UI france.html. Every page carries a country list linking the others.

Which area an object belongs to is recorded during the sweep — it is free, since the sweep is already chunked per area, and the GeoJSON carries no region field. German pages only appear on a build that sweeps all 16 Länder (PAPAMAP_COUNTRIES=dk must not publish an index claiming Germany has one Bundesland); every other country's page appears whenever that country's sweep is complete.

The URLs are fixed (the names are constants in pipeline/config.py), so they are listed by hand in web/sitemap.xml; tests/test_pages.py asserts that list matches the slugs the generator writes. PAPAMAP_PAGES_DIR moves the output elsewhere.

Play corners

Every feature carries a boolean play: true when the object also records an indoor place for the kid to play (kids_area:indoor or kids_area = yes|indoor|designated, leisure=indoor_play, or leisure=playground + indoor=yes). outdoor, no and limited are excluded, and an explicit kids_area:indoor=no overrules a bare kids_area=yes. The map draws it as a blue halo under the pin and the chip bar gains a filter that narrows to those places.

It is a badge, not a fourth status. A missing kids_area is silent across all ~13k pins and means nothing, so — unlike a grey changing_table:location pin — there is no "unknown" state to render and no call to action attached to it. The chip therefore starts off and subtracts, while the three status chips start on. Costs no extra Overpass query: the sweep already returns every tag on these objects. Measured on the DE+DK build of 17 Aug 2026: 828 objects pass the rule and 111 of them are already pins (48 accessible / 13 female_only / 50 unknown). Those are DE+DK figures and the 44-country sweep is larger, so the served numbers come from stats.json (local.play_tables, local.play_places) — the methods pages read them from there rather than repeating a number that moves every night.

Places to play

The other 701 of those DE+DK objects pass the same rule and carry no changing_table tag at all, so they are not pins and never could be — nobody has answered the first question about them. They get their own file, web/data/play_places.geojson, their own hollow-blue-ring layer and their own chip, off by default. changing_table=no places stay out: somebody did answer.

They are the best-targeted open questions on the map. A father with a toddler is going to an indoor playground or a café with a ball pit anyway, and while he is there he knows the answer. So the popup leads with the MapComplete link, the theme has a dad_play_place layer so that link lands on a selectable object, and the layer's first question is "does this place have a baby changing table?"

The sweep pays nothing for them: one union Overpass query per area returns both halves, and osm.split_sweep() sorts them apart by tag.

Leaderboard

A full build also appends one entry per day to web/data/history.json — [accessible, female_only, unknown] counts per region (the 16 Bundesländer, plus every swept country outside Germany under its own label) and per big city (the curated CITY_AREAS list, membership via one ids-only Overpass query per city) — and renders web/wickeltische/rangliste.html (German) plus leaderboard.html (English) from it. The tables rank the change of the answered share in percentage points against a snapshot at least a week back, never the absolute counts: levels measure mapping thoroughness, movement measures people answering the room question, and only the latter is an honest race. A same-date re-run replaces its history entry; partial builds write no history at all.

python -m pipeline.backfill 2026-07-17 2026-07-24 ... seeds past days from Overpass attic ([date:...]) queries so the page can show a real week-over-week delta from day one. It never overwrites a day that already exists. Attic queries over a whole Land run for minutes, so give it PAPAMAP_OVERPASS_QL_TIMEOUT=300 and a host without a 60 s connection cutoff (see docs/DEPLOY.md).

Every Overpass answer is checked for freshness (osm3s.timestamp_osm_base, PAPAMAP_OVERPASS_MAX_DATA_AGE_H, default 24 h): a mirror serving a frozen database is skipped, because a region quietly computed from months-old data is a data bug on the map and a fake mover on the leaderboard.

Serve locally

python3 -m http.server -d web 8000   # http://localhost:8000

Tests

pip install -r requirements-dev.txt   # adds pytest
pytest -v                  # pipeline (Python, offline — fixtures only)
node --test web/*.test.js  # frontend pure functions (needs Node.js)

Cron

0 2 * * * cd /path/to/papa-map && docker compose run --build --rm pipeline >> pipeline.log 2>&1

(02:00 since the 44-country sweep — projected ~105 min, and the ops mail is at 05:30.) (Matches the deploy in docs/DEPLOY.md — adjust if your clone lives elsewhere. Running the pipeline outside Docker works too; the venv variant is in the same file.)

An optional second cron line runs python -m pipeline.ops: an anomaly-gated ops mail (stale data, missing files, count drops and count jumps — a widened sweep is not mapping activity) plus a Monday all-clear digest with the week's grey→green transitions — see the "Ops mail" section in docs/DEPLOY.md. The same run rewrites a public ops page, /ops.html: the report plus last night's per-area build results, per-region counts and the daily history — aggregates only, no traffic numbers. A password-protected copy with Cloudflare's per-day request totals is opt-in, see there.

Deploy

Static files only — no API and no database. The site runs as a caddy:2-alpine container that bind-mounts web/, so git pull is the deploy; serving the directory with any static web server works just as well. Instructions: docs/DEPLOY.md.

Live at papamap.de (www.papamap.de and papamap.jakubwaller.eu redirect there).

Licences

Code: MIT (see LICENSE). Data: © OpenStreetMap contributors, ODbL — attribution details in web/methods.html.