When would you choose a hash map over a balanced binary search tree?

Choose a hash map when you only need lookups, inserts and deletes by key, and you want average constant time. Choose a balanced tree when you need ordering: range scans, nearest key, or iterating in sorted order, all in logarithmic time. Trees also give you predictable worst case behavior, while a hash map can degrade badly with poor hashing or stall on a resize at the wrong moment.

My default is a hash map, because most of the time I am doing point lookups by id and constant average time is hard to beat. The moment the requirement involves ordering, though, I switch. Anything like give me all events between two timestamps, or find the next key after this one, is a range query, and a hash map cannot do it without scanning everything. That wants a tree, or in practice a sorted structure with binary search. I also think about the tail. Hash lookups are constant on average, but a bad hash function or a rehash can spike, and inside a latency budget I care about that. On a leaderboard feature I built we started with a dictionary of user to score and then had to sort the whole thing on every read. Once we moved to a sorted set in Redis, reads became logarithmic and the endpoint stopped falling over at peak. Same data, different access pattern, different structure.