When would you use Django select_related or prefetch_related?

I select only relationships the response really consumes and test with realistic parent and child counts. Large prefetches can create large IN clauses and retain many objects, while broad join chains can transfer duplicate columns and lock more rows when combined with select_for_update(). Query count, total rows, payload construction time, and memory decide the choice. Sometimes an annotated projection or separate paged endpoint is more honest than either loader.