"Caching invalidation can be tricky, but for something like this it's actually easy"

The point is that you shouldn't need it at all. There's nothing to invalidate. You don't even need to cache it because there's nothing complex enough to cache. The full dataset is tiny. Adding caching might make it slower, because the cost of computing the data is less than the cost of a cache miss.

My point was that downloading the whole thing compared to using the API could be thought as a form of caching. Since it's with data that doesn't change much, the usual difficulty that comes with caching is not here.

From my limited understanding, not much which is why I think caching is safe in that case.