Andrew Dunstan <andrew@dunslane.net> writes:
> Fixed
Now that we got past the hard failures, we can see that the test
falls over with (some?) non-default encodings, as for instance
here:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=prion&dt=2022-03-22%2020%3A23%3A13
I can replicate that by running the test under LANG=en_US.iso885915.
What I think is happening is:
(1) Rather unwisely, the relevant InvokeNamespaceSearchHook calls
appear in recomputeNamespacePath. That means that their timing
depends heavily on accidents of caching.
(2) If we decide that we need an encoding conversion to talk to
the client, there'll be a lookup for the conversion function
early during session startup. That will cause the namespace
search path to get computed then, before the test module has been
loaded and certainly before the audit GUC has been turned on.
(3) At the point where the test expects some audit notices
to come out, nothing happens because the search path is
already validated.
I'm inclined to think that (1) is a seriously bad idea,
not only because of this instability, but because
(a) the namespace cache logic is unlikely to cause the search-path
cache to get invalidated when something happens that might cause an
OAT hook to wish to change its decision, and
(b) this placement means that the hook is invoked during cache loading
operations that are likely to be super-sensitive to any additional
catalog accesses a hook might wish to do. (I await the results of the
CLOBBER_CACHE_ALWAYS animals with trepidation.)
Now, if our attitude to the OAT hooks is that we are going to
sprinkle some at random and whether they are useful is someone
else's problem, then maybe these are not interesting concerns.
regards, tom lane
