Hi,
On 2020-10-06 15:31:16 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2020-10-03 08:40:31 +0200, Peter Eisentraut wrote:
> >> Since we have ADD PRIMARY KEY USING INDEX, we can declare a primary key for
> >> an existing index. So this doesn't have to affect the low-level early
> >> bootstrapping. The attached patch adds those commands manually. Another
> >> option might be to have the catalog generation Perl tooling create those
> >> declarations automatically from some marker in the catalog files. That
> >> would also allow declaring unique constraints for the unique indexes that
> >> don't end up being the primary key.
>
> > Hm. What prevents us from declaring the pkey during bootstrap? I don't
> > at all like adding yet another place that needs manual editing when
> > doing DDL changes.
>
> We don't add new catalogs often, so the cost-benefit ratio of automating
> this looks pretty poor.
True, we don't create new ones that often. Still think that distributing
such setup over fewer places is good. And it's not like there's only a
handful of pkeys to start with. To me it makes more sense to add a
DECLARE_PRIMARY_KEY in indexing.h, replacing the relevant
DECLARE_UNIQUE_INDEX stanzas. Or, possibly preferrable, do it as part of
the CATALOG() directly - which'd avoid needing the index statement in
the first place.
The Catalog.pm part is trivial. It's abit harder to implement actually
creating the constraint - but even the hard route of adding a new field
to Boot_DeclareIndexStmt or such wouldn't be particularly hard.
> One thing I'm wondering about though is whether Peter's implementation
> ends up with desirable pg_depend contents for the pkey constraints.
> Probably we want them to just be pinned and not have any explicit
> dependencies on the associated indexes, but I haven't thought hard
> about it.
That sounds right to me.
Wonder whether it's not time to move the DECLARE bits from indexing.h to
the individual catalogs, independent of what we're discussing here. With
todays Catalog.pm there's really not much point in keeping them
separate, imo.
Greetings,
Andres Freund