Re: VACUUM can finish an interrupted nbtree page split -- is that okay?

On Mon, May 13, 2019 at 9:09 PM Peter Geoghegan <pg@bowt.ie> wrote:
> Even when that happens, the index is already considered corrupt by
> VACUUM, so the same VACUUM process that could in theory be adversely
> affected by removing the half-dead internal page check will complain
> about the page when it gets to it later on -- the user will be told to
> REINDEX. And even then, we will never actually get to apply the check
> that I propose to remove, since we're already checking the leaf page
> sibling of the leaf-level target -- the leaf-level test that was added
> by efada2b8e92 was clearly necessary. But it was also sufficient (no
> equivalent internal half-dead right sibling test is needed): a 9.3-era
> half-dead internal page cannot have more than one child, which must be
> undergoing deletion as well.

Actually, now that I look back at how page deletion worked 5+ years
ago, I realize that I have this slightly wrong: the leaf level check
is not sufficient to figure out if the parent's right sibling is
pending deletion (which is represented explicitly as a half-dead
internal page prior to 9.4). All the same, I'm going to push ahead
with this patch. Bugfix commit efada2b8e92 was always about a bug in
9.4 -- it had nothing to do with 9.3. And, in the unlikely event that
there is a problem on a pg_upgrade'd 9.3 -> 12 database that happens
to have half-dead internal pages, we'll still get a useful, correct
error from VACUUM one way or another. It might be slightly less
friendly as error messages about corruption go, but that seems
acceptable to me.

-- 
Peter Geoghegan