Re: Bug in pg_dump --table and --exclude-table for declarative partition table handling.

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Thu, May 11, 2017 at 9:02 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> You could argue that it would be better for pg_dump to emit something
>> like
>>
>> CREATE TABLE c (...);
>> ALTER TABLE c INHERIT p;
>>
>> so that if p isn't around, at least c gets created. And I think it
>> *would* be better, probably. But if that isn't a new feature then
>> I don't know what is. And partitioning really ought to track the
>> behavior of inheritance here.

> Hmm, I think that'd actually be worse, because it would break the use
> case where you want to restore the old contents of one particular
> inheritance child. So you drop that child (but not the parent or any
> other children) and then do a selective restore of that one child.
> Right now that works fine, but it'll fail with an error if we try to
> create the already-extant parent.

Uh ... what in that is creating the already-extant parent? What
I'm envisioning is simply that the TOC entry for the child table
contains those two statements rather than "CREATE TABLE c (...)
INHERITS (p)". The equivalent thing for the partitioned case is
to use a separate ATTACH PARTITION command rather than naming the
parent immediately in the child's CREATE TABLE. This is independent
of the question of how --table and friends ought to behave.

> I think one answer to the original complaint might be to add a new
> flag to pg_dump, something like --recursive-selection, maybe -r for
> short, which makes --table, --exclude-table, and --exclude-table-data
> cascade to inheritance descendents.

Yeah, you could do it like that. Another way to do it would be to
create variants of all the selection switches, along the lines of
"--table-all=foo" meaning "foo plus its children". Then you could
have some switches recursing and others not within the same command.
But maybe that's more flexibility than needed ... and I'm having a
hard time coming up with nice switch names, anyway.

Anyway, I'm still of the opinion that it's fine to leave this as a
future feature. If we've gotten away without it this long for
inherited tables, it's unlikely to be critical for partitioned tables.

regards, tom lane