Skip to content

1183 — Train CSV station disambiguation

1. Problem

The train CSV trip format ships station names (origin_name, destination_name). Same-name stations exist in multiple countries (BERNE, CH vs BERNE, DE), so the CSV-time resolver in ProfessionalTravelTrainModuleHandler.enrich_csv_row cannot disambiguate cross-country collisions from names alone.

The earlier approach defaulted the missing country to CH. That silently mis-resolved every non-Swiss station and violated the no-silent-fallbacks rule. This iteration removes the default and requires the country instead — made viable by sourcing the seed from a dataset that carries country natively.

2. Location seed: trainline-eu stations.csv

The train location seed is built from the open trainline-eu dataset (https://github.com/trainline-eu/stations), which ships an ISO-2 country per station — so country codes come from the source, not a backfill.

  • Builder: backend/scripts/build_train_seed_from_trainline.py (stdlib csv only; LF terminators; ≤40-line functions).
  • Input backend/stations.csv (;-delimited, gitignored) → output backend/seed_data/seed_travel_location_train.csv (the 10-column comma schema ReferenceDataCSVProvider ingests).
  • Kept rows: is_suggestable=t, is_airport≠t, non-empty latitude/longitude (NOT NULL in locations), non-empty country. → 51,299 stations.
  • continent/municipality/iata_code/airport_size are absent from the source and stay blank (all optional for trains); keywords mirrors name for station search.

Re-upload via the backoffice train station reference slot (data_entry_type_id=21, not plane=20).

3. Required country_code in the trip resolver

enrich_csv_row now rejects any train CSV row that lacks a {role}_country_code for an endpoint without a precomputed {role}_natural_key — before any station lookup. There is no CH default.

  • app/modules/professional_travel/schemas.py::enrich_csv_row — missing country_code → row error Missing {role}_country_code.
  • app/services/location_service.py::resolve_train_station_for_csv — country_code is now a required parameter (no default_country_code="CH").
  • UI/API entries are unaffected: they carry *_natural_key from the station autocomplete and skip the resolver branch entirely.

location_repo.search_locations keeps its CH-first autocomplete ranking — that is UI ordering, not an ingestion default, and is out of scope here.

4. Tests

  • tests/unit/services/data_ingestion/test_train_enrich_csv_row.py — new. Asserts a row missing origin/destination country_code is rejected and never queries the DB (sentinel session). Fast, no Postgres.
  • tests/integration/.../test_travel_pg.py::test_train_csv_resolves_station_by_required_country_code — renamed from the CH-default-override test; pins the CH-vs-DE Berne collision, asserting destination_country_code=DE resolves to the German station.

5. Backfill risk (accepted)

Legacy DataEntry rows ingested under the old CH-default resolver may carry natural_keys pointing at the Swiss station. Per project policy ("No DB backfill until v1.x" — v0.x drops the DB between deploys) we accept the drift and do not migrate historical rows.

6. Verification

  1. cd backend && uv run python scripts/build_train_seed_from_trainline.py → wrote 51299 train stations.
  2. uv run pytest tests/unit/services/data_ingestion/test_train_enrich_csv_row.py
  3. uv run pytest tests/integration/services/data_ingestion/test_travel_pg.py -k train
  4. make type-check

7. Files

  • backend/scripts/build_train_seed_from_trainline.py — new (replaces the retired add_country_codes_to_train_csvs.py).
  • backend/app/modules/professional_travel/schemas.py — require country_code.
  • backend/app/services/location_service.py — drop CH default param.
  • backend/tests/unit/services/data_ingestion/test_train_enrich_csv_row.py — new.
  • backend/tests/integration/services/data_ingestion/test_travel_pg.py — test rename/rescope.
  • docs/src/implementation-plans/1183-train-csv-country-code.md — this plan.