Re: Extension Packaging

On Thu, Apr 21, 2011 at 4:16 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> "David E. Wheeler" <david(at)kineticode(dot)com> writes:

>> * Another, minor point: If I release a new version with no changes to the code (as I've done today, just changing the make stuff and documentation), it's kind of annoying that I'd need to have a migration script from the old version to the new version that's…empty. But I dunno, maybe not such a big deal. It's useful to have it there with a comment in it: “No changes.”
>
> If you did not actually change the contents of the install script, you
> should not change its version number either.

Sorry, I'm not entirely convinced. If I release an extension 1.0, then
find a bug in the C code and fix it in 1.0.1, arguably "make install"
will put the .so in the right place and the 1.0.1 code will be picked
up by new sessions. But pg_extension still knows 1.0 as extension
version, and ALTER EXTENSION ... UPGRADE fails because no update path
is knows.

There is also a dangerous asymmetry: If I'm not mistaken the library
.so has no version number, so there can be only one version in the
system: an update changing code and sql requires ALTER EXTENSION to be
run as soon as possible, or some sql function from the old extension
may try to call non-existing functions in the library - or worse the
wrong ones or with wrong parameters. OTOH library-only changes don't
strictly require ALTER EXTENSION - and trying to issue the command
would fail if no path to the default version is available (leaving the
admin puzzled about whether he installed the upgrade properly).

Is an empty upgrade file the only way to get the extension metadata
right? I wouldn't find it particularly bad, but better have it that
have library-metadata mismatches or inconsistencies in the upgrade
procedures I think.

-- Daniele