On Mon, Oct 10, 2011 at 2:52 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> But there's a bigger problem: it seems to me that we have an
>> inconsistency between what happens when you create an extension from
>> scratch and when you upgrade it from unpackaged. Both pg_buffercache
>> and pg_stat_statements just do this in the "upgrade from unpackaged"
>> case:
>
>> ALTER EXTENSION <ext-name> ADD view <view-name>;
>
>> They do *not* add the type and the array type. But when the "1.0"
>> script is run, the type and array type end up belonging to the
>> extension. This seems bad.
>
> Hmm, yeah, we need to make those consistent.
>
> The underlying issue here is whether objects dependent on an extension
> member should have direct dependencies on the extension too, and if not,
> how do we prevent that? The recordDependencyOnCurrentExtension calls
> don't have enough information to know what to do, I think.
After looking at this code, it seems that we've generally made that
the caller's problem - e.g. in heap_create_with_catalog(), we skip
recordDependencyOnCurrentExtension() if we're dealing with a composite
type. So I think the fix here is just to move the
recordDependencyOnCurrentExtension() call in pg_type.c inside the
if-block that precedes it, as in the attached patch.
Of course, this won't fix any damage already done, but it seems like
the right thing going forward.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company