On Wed, Apr 29, 2020 at 1:40 PM Peter Geoghegan <pg@bowt.ie> wrote:
> I'm not sure how low the costs would be, but at least we'd only have
> to do it once per already-deleted page (i.e. only the first time a
> VACUUM operation found _bt_page_eligible_for_recycling() returned true
> for the page and marked it recycled in a crash safe manner). That
> design would be quite a lot simpler, because it expresses the problem
> in terms that make sense to the nbtree code. _bt_getbuf() should not
> have to make a distinction between "possibly recycled" versus
> "definitely recycled".
As a bonus, we now have an easy correctness cross-check: if some
random piece of nbtree code lands on a page (having followed a
downlink or sibling link) that is marked recycled, then clearly
something is very wrong -- throw a "can't happen" error.
This would be especially useful in places like _bt_readpage(), I
suppose. Think of extreme cases like cursors, which can have a scan
that remembers a block number of a leaf page, that only actually gets
accessed hours or days later (for whatever reason). If that code was
buggy in some way, we might have a hope of figuring it out at some
point with this cross-check.
--
Peter Geoghegan
