I agree with the basic idea in your blog post -- that interfaces should be defined by the consumer, and be small -- but I think it's a bit dogmatic: "it’s wrong". :-)
For one thing, it's easy to work around the issue of "pasting every single function signature into this test file, and making it panic("not implemented") or something". You can simply embed the DB interface in your test struct, and only override the method(s) you need. For example, let's say the DB interface defines 10 methods, but our function under test only need one, GetThings(). We do this:
type testDB struct {
DB // embedded field named DB, so testDB implements DB
things []Thing
}
// but override GetThings method
func (d *testDB) GetThings() ([]Thing, error) {
return d.things, nil
}
The embedded DB field, which is a field named DB of type DB and is nil, pulls in all the methods of that type so testDB implements the DB interface. When you construct a testDB you just leave the DB field nil -- if any of the not-explicitly-defined methods are called, they'll panic with a nil pointer, but that's okay, they would explicitly panic in your test implementation anyway.
The other thing is simply pragmatism: at some point it gets tedious to define all these tiny interfaces. Do you do it at the handler level, where each db sub-interface only has 1 or 2 methods? This blows out the number of interfaces. Or do you do it at the server level, where the db interface has all the methods (what you're opposed to)? Or something in between: have server "sections" grouped by logical feature, such as authDB, projectDB, and so on. Each might have 5-10 methods, and to test each sub-server you'd either do what I've suggested above with an embedded interface field, or just implement a single in-memory test db for each db type.
I do like what you said about interfaces obfuscating the code (it becomes more abstract and hard to follow and navigate), and that testing on the real db if you can is a good idea. Creating a new PostgreSQL db for every test seems like it'd be very slow and unnecessary, though. Why not just one db for each test run?