If we use unordered_set<T> (T is custom type, not simple type), it's quite likely that we need define custom operator== and hasher, which is kind of cumbersome if we have lots of different T in a big project. Furthermore how to choose a good hash function is a big issue.

But if we use unordered_set<T*> (including smart pointer), we can omit those definitions. Especially I find in rocksdb and mysql-server codebase, lots of unordered_set<T*> are used instead of unordered_set<T>, whereas in mongodb and scylladb, lots of unordered_set<T> defined. This situation gives me impression that rocksdb developers seems intentionally to avoid using unordered_set<T> ?

So I'm a little confused when to use unordered_set<T> and when use unorderd_set<T*> ?