Re: Fix REPACK with WITHOUT OVERLAPS replica identity indexes

On 2026-May-13, Chao Li wrote:

> After some investigation, I think I see the mismatch:
>
> * get_relation_info(): non-ordered GiST cannot provide sort order.
> That is expected.
> * copy_table_data() only uses plan_cluster_use_sort() for btree. For
> any other clusterable index, it sets use_sort = false and does a raw
> index scan.
> * The docs say REPACK can re-sort using index scan “if the index is a
> b-tree” or seqscan+sort, which does not describe what the code
> actually does for GiST.
>
> I am not sure whether we should change the behavior in PG19. Alvaro
> may have a better idea about that. But I agree that we can at least
> clarify the code comment and documentation. The attached patch
> attempts to do that.

I am sure that we should not change the behavior for 19. We could
change it in 20, perhaps. But I think the really gross problem is that
somehow I made the documentation say exactly the opposite of truth, in
the rework for commit ac58465e0618. I fixed that just now with by
pushing 07aef52eba3c, which includes your (Chao's) proposed change; but
also, while reading and rereading the existing docs and your patch, it
struck me that the preexisting text was not super nice either: there was
one introductory text, then some technical details, then some more
discussion, then more technicalities. That was weird and probably just
historical accident. I reordered the text and made some other wording
changes. I *think* result is reasonable, or at least it contains no
egregious errors; but maybe it could be further improved, if anybody is
interested.

--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/
"El Maquinismo fue proscrito so pena de cosquilleo hasta la muerte"
(Ijon Tichy en Viajes, Stanislaw Lem)