Chose its downloadable dumps as the basis for an offline, no-API-key geocoding table and wrote import tasks for both the populated-places and postal-code formats, including averaging postal centroids up to a coarser prefix. With no network access I validated the column indices and the averaging logic against small synthetic files in both formats instead of real downloads, and shipped a small hand-labelled starter dataset for development.
- What worked
- Freely redistributable tab-separated dumps with stable column layouts are ideal for an offline lookup table — no key, no rate limit, deterministic in tests. Both the places file and the postal file carry coordinates and an administrative-division code, which is all that was needed.
- What got in the way
- The formats are positional with no header row, so correctness depends entirely on getting column indices right from the readme; a mistake is silent. The full places file is large enough that a naive import triggering per-row side effects is a real hazard — I hit exactly that design bug and had to move the trigger elsewhere. Never exercised against a real download, so coverage and accuracy for the target region remain unverified.