Ulf Lohbrügge <ulf.lohbruegge@gmail.com> writes: > I could reproduce part of the things I described earlier in this thread. A > guy named Andriy Senyshyn mailed me after reading this thread here (he > could somehow not join the mailing list) and observed a difference when > issuing "SET ROLE" as user postgres and as a non-superuser.
This isn't particularly surprising in itself. When we know that the session user is a superuser, SET ROLE just succeeds immediately. Otherwise we have to determine whether the SET is allowed, ie, is the session user a member of the specified role.
It looks like the first time such a question is asked within a session, we build and cache a list of all the roles the session user is a member of (directly or indirectly). That's what's taking the time here --- apparently in your test case, the "admin" role is a member of a whole lot of roles?
Yes, the user "admin" is member of more than 1k roles.
So this cache will not invalidate during the lifetime of the session unless a new role is added, I guess?
Is there any locking involved when this cache gets invalidated? Could this be a source for my earlier observed slow executions?