Re: BUG #18240: Undefined behaviour in cash_mul_flt8() and friends

On Fri, Jan 12, 2024 at 10:50:25PM -0500, Jonathan S. Katz wrote:
> On 1/12/24 11:33 AM, Tom Lane wrote:
>> What'd have to happen is that people would have to upgrade to a
>> version of btree_gin/btree_gist that deletes its money support
>> before they could pg_upgrade into a core version that lacks money.
>> So we'd have to ship that version at least one major release before
>> nuking the core support.
>
> Hm -- what about people who skip versions (e.g. 16 => 19)? Would they have
> to stop at 18 first to perform the upgrade? And does this only affect people
> who use btree_gist/gin for moeny, or all btree_gist/gin users?

That's exactly why I am worried about a schedule being too aggressive
here. Less runs of pg_upgrade means less headaches and less downtime.
I'd guess that we should have at least 3~4 years between a deprecation
in the docs and the actual removal. It seems like this is what we
should do now on HEAD, aiming at a complete removal in 21~22 when 17
gets out of support (even if pg_upgrade support is down to 10 years).

>> v19: remove it from core
>>
>> Note that in v18, people could still use money even in
>> btree_gin/btree_gist, just by installing a non-default extension
>> version. So their C code for money would have to stay.
>
> Yeah, but for what we support directly in PostgreSQL, we will have made best
> effort. If someone _really_ wants to do something with "money" in the way
> you describe above, then they're committed to it.

By the way, let me know if you have any objections about the patch I
have posted upthread that reduces the dependency to money in the main
regression test suite, replacing things with varbit while keeping
coverage the same. I have a few second thoughts about the change in
create_table, so I'm tempted to leave that part out for now, switching
the rest.
--
Michael