On Fri, Jan 12, 2024 at 9:44 AM Robert Haas <robertmh...@gmail.com> wrote: > Looking through the git history, I see that this behavior seems to > date back to 44fa84881fff4529d68e2437a58ad2c906af5805 which introduced > lazy_scan_noprune(). The comments don't explain the reasoning, but my > guess is that it was just an accident. It's not entirely evident to me > whether there might ever be good reasons to update the freespace map > for a page where we haven't freed up any space -- after all, the free > space map isn't crash-safe, so you could always postulate that > updating it will correct an existing inaccuracy. But I really doubt > that there's any good reason for lazy_scan_prune() and > lazy_scan_noprune() to have different ideas about whether to update > the FSM or not, especially in an obscure corner case like this.
Why do you think that lazy_scan_prune() and lazy_scan_noprune() have different ideas about whether to update the FSM or not? Barring certain failsafe edge cases, we always call PageGetHeapFreeSpace() exactly once for each scanned page. While it's true that we won't always do that in the first heap pass (in cases where VACUUM has indexes), that's only because we expect to do it in the second heap pass instead -- since we only want to do it once. It's true that lazy_scan_noprune unconditionally calls PageGetHeapFreeSpace() when vacrel->nindexes == 0. But that's the same behavior as lazy_scan_prune when vacrel-> nindexes == 0. In both cases we know that there won't be any second heap pass, and so in both cases we always call PageGetHeapFreeSpace() in the first heap pass. It's just that it's a bit harder to see that in the lazy_scan_prune case. No? -- Peter Geoghegan
