On Tue, Jan 10, 2023 at 06:50:31PM -0500, Tom Lane wrote:
> With the proposed % feature, if foo--%--3.0.sql exists then the
> system will invoke it and expect the end result to be a valid
> 3.0 installation, whether or not the script actually has any
> ability to do a downgrade.
It is sane, for the system to expect the end result
to be a valid 3.0 installation if no exception is
thrown by the script itself.
If we ship foo--%--3.0.sql we must have been taken care of
protecting from unsupported downgrades/upgrades (and we did,
having the script throw an exception if anything is unexpected).
> (I suppose you could add code to look at pg_extension.extversion,
We actually added code looking at our own version-extracting
function (which existed since before PostgreSQL added support
for extensions). Is the function not there ? Raise an exception.
Is the function not owned by the extension ? Raise an exception.
In other cases -> trust the output of that function to tell what
version we're coming from, throw an exception if upgrade to the
target version is unsupported.
> to add such code is that really better than having a number of
> one-liner scripts implementing the same check declaratively?)
Yes, it is, because no matter how many one-liner scripts we install
(we currently install 87 * 6 such scripts, we always end up missing
some of them upon releasing a new bug-fix release in stable branches.
> almost certainly you are
> going to end with an extension containing some leftover 4.0
> objects, some 3.0 objects, and maybe some objects with properties
> that don't exactly match either 3.0 or 4.0.
This is already possible, depending on WHO writes those upgrade
scripts. This proposal just gives more expressiveness power to
those script authors.
> So I really think this is a case of "be careful what you ask
> for, you might get it". Even if PostGIS is willing to put in
> the amount of infrastructure legwork needed to make such a
> design bulletproof, I'm quite sure nobody else will manage
> to use such a thing successfully.
This is why I initially made this something to be explicitly enabled
by the .control file, which I can do again if it feels safer for you.
> I'd rather spend our
> development effort on a feature that has more than one use-case.
Use case is any extension willing to support more than a single stable
branch while still allowing upgrading from newer-stable-bugfix-release
to older-feature-release (ie: 3.2.10 -> 3.4.0 ). Does not seem so
uncommon to me, for a big project. Maybe there aren't enough big
extension-based projects out there ?
--strk;
Libre GIS consultant/developer
https://strk.kbt.io/services.html