Robert Haas <robertmhaas@gmail.com> writes:
> On Thu, Apr 27, 2017 at 5:22 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> How so? Shouldn't the indexscan go back and mark such tuples dead in
>> the index, such that they'd be visited this way only once? If that's
>> not happening, maybe we should try to fix it.
> Hmm. Actually, I think the scenario I saw was where there was a large
> number of tuples at the end of the index that weren't dead yet due to
> an old snapshot held open. That index was being scanned by lots of
> short-running queries. Those queries executed just fine, but they
> took a long to plan because they had to step over all of the dead
> tuples in the index one by one.
But that was the scenario that we intended to fix by changing to
SnapshotDirty, no? Or I guess not quite, because
dead-but-still-visible-to-somebody tuples are rejected by SnapshotDirty.
Maybe we need another type of snapshot that would accept any
non-vacuumable tuple. I really don't want SnapshotAny semantics here,
but a tuple that was live more recently than the xmin horizon seems
like it's acceptable enough. HeapTupleSatisfiesVacuum already
implements the right behavior, but we don't have a Snapshot-style
interface for it.
regards, tom lane