On Sun, Aug 23, 2009 at 7:34 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
> Greg Stark <gsstark@mit.edu> writes:
>> On Sun, Aug 23, 2009 at 7:00 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
>>> But we don't use that while examining individual tuples, do we?
>
>> We don't use the visibility map itself but we *do* use the page
>> header's all visible bit. On a sequential scan we skip the visibility
>> check for tuples on the page if the page header bit is set.
>
> Urgh. I bet that's where the problem is then. Some path is failing
> to clear that bit, or maybe there's a race condition that allows it
> to become set incorrectly (ie right after somebody else adds a tuple).
Well it only gets set in vacuum when we have the vacuum cleanout lock
so I can't see it getting set incorrectly unless there's a simple
logic error in vacuum. I just scanned it and it looks right to me.
I can imagine it failing to be unset. Or being unset but allowing a
race condition window before it's unset when it's wrong. That doesn't
explain why it the HEAP_ONLY_TUPLE flag would be missing though.
Also, vacuum logs a warning if the flag is found to be set and
shouldn't be. I suppose that could easily be missed though if it's
autovacuum that ran it. Also, the flag might fail to be cleared but
then be cleared later by another operation. I wonder how much damage
it can do while it's wrong.
--
greg
http://mit.edu/~gsstark/resume.pdf
