Andy Fan <zhihui.fan1213@gmail.com> writes:
> On Wed, Nov 25, 2020 at 9:40 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> What exactly would be the value of that?
>> ...
> I agree with this, but I don't think there is no value in my suggestion
> unless I missed something. Per my current understanding, the code
> is too easy to make the datadir incompatible with binary, which requires
> user to do some extra work and my suggestion is to avoid that and
> it is the value.
It'd be fairly pointless to worry about this so far as ordinary users are
concerned, who are only going to encounter the situation in connection
with a major-version upgrade. There is no way that the only catalog
incompatibility they'd face is an addition or removal of a field in some
query node. In practice, a major-version upgrade is also going to
involve things like these:
* addition (or, sometimes, removal) of entire system catalogs
* addition, removal, or redefinition of columns within an existing catalog
* addition or removal of standard entries in system catalogs
* restructuring of stored rules/expressions in ways that are more
complicated than simple addition/removal of fields
The second of those, in particular, is quite fatal to any idea of
making a version-N backend executable work with version-not-N
catalogs. Catalog rowtypes are wired into the C code at a pretty
basic level.
We already sweat a great deal to make user table contents be upwards
compatible across major versions. I think that trying to take on
some similar guarantee with respect to system catalog contents would
serve mostly to waste a lot of developer manpower that can be put to
much better uses.
regards, tom lane