There's a problem with the current char by char approach cause of the nature of the usernames we use for load-test. Since they need to be predictable they all share a common prefix. This means that calls to AutoCompleteUsersInChannel might stress the DB more than on an instance with very different usernames.
One way to fix this would be to start the search from the first character that is not common to all users (probably the numeric ID).
RIght now to mention testuser100 we would search for:
This could give potentially thousands of results, unnecessarily and unrealistically stressing the DB.
What we might want to do instead is search for: