In the fast case the 'Felhasználók%' part is known at query planning time, so it can be a prefix search.
In the slow case, the planner doesn't know what that value will be, it could be something that starts with '%' for example.
First of all, it CANNOT start with '%'. This is a fact and this fact can be determined by analyzing the query. Something that the query planner should do, right?
Second argument: the same query is also slow with the ^@ operator...
Oh I see, the query planner does not know that there will be no % characters in file and folder names.
But what is the solution then? It just seems wrong that I can speed up a query 1000 times by replacing it with a nested loop in a pl/sql function :(