On Mon, Oct 23, 2023 at 6:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > Well, in practice "indexUnchanged = true" means "do bottom-up deletion
> > if it's the only way to avoid a page split". The justification is that
> > the incoming tuple is "logically unchanged" (actually it's more
> > complicated than that, but that's our starting point).
>
> But doesn't the need for a non-HOT update show that the tuple *was*
> changed --- in index-relevant columns, even? Maybe I'm still not
> understanding exactly what condition we're detecting.
Not necessarily. For one thing you might need to do a non-HOT update
purely because there isn't enough free space to fit the successor
version on the original heap page. In general, even a 100% HOT-safe
UPDATE statement might not be able to perform HOT updates.
As I said earlier, the way that we deal with partial indexes doesn't
really make too much sense if you try to shoehorn it into an abstract
definition. It makes a lot more sense when seen in the context of a
workload with a partial index, and shown by a test case from Marko
Tiikkaja:
https://www.postgresql.org/message-id/flat/CAL9smLC%3DSxYiN7yZ4HDyk0RnZyXoP2vaHD-Vg1JskOEHyhMXug%40mail.gmail.com#e79eca5922789de828314e296fdcb82d
I freely admit that this general approach is non-modular, even ugly.
--
Peter Geoghegan