On Fri, Nov 13, 2020 at 10:03:43AM -0500, Stephen Frost wrote: > * Magnus Hagander (mag...@hagander.net) wrote: > > On Thu, Nov 12, 2020 at 11:28 PM Tom Lane <t...@sss.pgh.pa.us> wrote: > > > > The changes to the contrib modules appear to be incomplete in some ways. > > > > In cube, hstore, and seg, there are no changes to the extension > > > > scripts to remove the operators. All you're doing is changing the C > > > > code to no longer recognize the strategy, but that doesn't explain what > > > > will happen if the operator is still used. In intarray, by contrast, > > > > you're editing an existing extension script, but that should be done by > > > > an upgrade script instead. > > > > > > In the contrib modules, I'm afraid what you gotta do is remove the > > > SQL operator definitions but leave the opclass code support in place. > > > That's because there's no guarantee that users will update the extension's > > > SQL version immediately, so a v14 build of the .so might still be used > > > with the old SQL definitions. It's not clear how much window we need > > > give for people to do that update, but I don't think "zero" is an > > > acceptable answer. > > > > Based on my experience from the field, the answer is "never". > > > > As in, most people have no idea they are even *supposed* to do such an > > upgrade, so they don't do it. Until we solve that problem, I think > > we're basically stuck with keeping them "forever". (and even if/when > > we do, "zero" is probably not going to cut it, no) > > Yeah, this is a serious problem and one that we should figure out a way > to fix or at least improve on- maybe by having pg_upgrade say something > about extensions that could/should be upgraded..?
I think what's needed are: 1) a way to *warn* users about deprecation. CREATE EXTENSION might give an elog(WARNING), but it's probably not enough. It only happens once, and if it's in pg_restore/pg_upgrade, it be wrapped by vendor upgrade scripts. It needs to be more like first function call in every session. Or more likely, put it in documentation for 10 years. 2) a way to *enforce* it, like making CREATE EXTENSION fail when run against an excessively old server, including by pg_restore/pg_upgrade (which ought to also handle it in --check). Are there any contrib for which (1) is done and we're anywhere near doing (2) ? -- Justin
