Robert Haas <robertmhaas@gmail.com> writes:
> On Fri, Jun 17, 2011 at 1:22 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Yeah. After mulling it for awhile, what about this idea: we could
>> redefine SnapshotNow as a snapshot type that includes a list of
>> transactions-in-progress, somewhat like an MVCC snapshot, but we don't
>> fill that list from the PGPROC array. Instead, while running a scan
>> with SnapshotNow, anytime we determine that a particular XID is
>> still-in-progress, we add that XID to the snapshot's list.
> I think that something like that might possibly work, but what if the
> XID array overflows?
Well, you repalloc it bigger. In either this idea or yours below,
I fear SnapshotNow snaps will have to become dynamically-allocated
structures instead of being simple references to a shared constant
object. (This is because we can sometimes do a SnapshotNow scan when
another one is already in progress, and we couldn't let the inner one
change the outer one's state.) That's not really a performance problem;
one more palloc to do a catalog scan is a non-issue. But it is likely
to be a large notational change compared to what we've got now.
> A while back I proposed the idea of a "lazy" snapshot, by which I had
> in mind something similar to what you are suggesting but different in
> detail. Initially, when asked to acquire a snapshot, the snapshot
> manager acknowledges having taken one but does not actually do any
> work. As long as it sees only XIDs that either precede the oldest XID
> still running anywhere in the cluster, or have aborted, it can provide
> answers that are 100% correct without any further data. If it ever
> sees a newer, non-aborted XID then it goes and really gets an MVCC
> snapshot at that point, which it can uses from that point onward. I
> think that it might be possible to make such a system work even for
> MVCC snapshots generally, but even if not, it might be sufficient for
> this purpose. Unlike your approach, it would avoid both the "see no
> rows" and the "see multiple rows" cases, which might be thought an
> advantage.
Hmm, yeah, I think this idea is probably better than mine, just because
of the less dubious semantics. I don't see how you'd make it work for
generic MVCC scans, because the behavior will be "the database state as
of some hard-to-predict time after the scan starts", which is not what
we want for MVCC. But it ought to be fine for SnapshotNow.
regards, tom lane
