On Fri, Feb 14, 2020 at 11:35 AM Michael Paquier <michael@paquier.xyz> wrote:
On Thu, Feb 13, 2020 at 09:05:01PM +0530, Ashutosh Bapat wrote: > That looks odd. Other sessions are able to see temporary tables of a given > session because they are stored in the same catalog which is accessible to > all the sessions. But ideally, a temporary table should be visible only to > the session which created it (GTT is an exception). So LOCK and DROP table > should not succeed.
One thing that we need to consider is if there are applications which take advantage of LOCK allowed on temp relations from other backends or not. One downside is that if one backend takes a lock on a temp table from a different session, then this other session would not completely shut down (still report the shutdown to the client), and would remain blocked during the temp schema cleanup until the transaction of the session locking the temp relation commits. This blocks access to one connection slot, still we are talking about an operation where the owner of the temp schema wants to do the lock.
That might be disastrous if happens by accident eating up most of the available connection slots.
Whatever the user wants to achieve using LOCK [temp] TABLE of other session, I guess can be achieved by other means or can be shown to have disastrous effect. So that kind of usage pattern would better be forced to change.
> Thinking from a different perspective, DROP TABLE being able to drop a > temporary table seems a good tool in case a temporary table is left behind > by a finished session. But that doesn't seem like a good reason to have it > and I don't see much use of LOCK TABLE there.
Yep. Robert had actually this argument with DROP SCHEMA pg_temp not so long ago with me.
DROP SCHEMA might be better for mass cleanup. That actually makes DROP [other session temp] TABLE useless.