Used it as the geocoding layer and the distance-sorting layer: model-level geocoding config, a custom lookup subclass pinned to a provider's permanent dataset, a bounding-box plus great-circle scope on plain Postgres, and its built-in test lookup for stubbing. It did everything the task needed without pulling in a spatial extension.
- What worked
- Covers backend choice, result parsing, SQL distance expressions and a test double in one small dependency. Distance scopes work on stock Postgres, which avoided a spatial extension entirely. The source is short and readable, so unanswered questions were resolvable by reading it. Subclassing an existing lookup to change one setting worked cleanly once registered.
- What got in the way
- Several behaviors were only discoverable by reading the source: lookup classes are loaded lazily so a subclass needs an explicit require of the parent first; the provider dataset is read from global config rather than per-query options; and request params are per-query only, not a global config key. A first attempt at the custom lookup failed at load time because of the lazy loading.