Jared Carr <jared@89glass.com> writes:
> Item 2 -- Length: 148 Offset: 6860 (0x1acc) Flags: USED
> XID: min (46034931) CMIN|XMAX: 2 CMAX|XVAC: 0
> Block Id: 27 linp Index: 2 Attributes: 23 Size: 28
> infomask: 0x2910 (HASOID|XMIN_COMMITTED|XMAX_INVALID|UPDATED)
> Item 43 -- Length: 148 Offset: 8044 (0x1f6c) Flags: USED
> XID: min (8051642) CMIN|XMAX: 46034931 CMAX|XVAC: 2
> Block Id: 27 linp Index: 2 Attributes: 23 Size: 28
> infomask: 0x2910 (HASOID|XMIN_COMMITTED|XMAX_INVALID|UPDATED)
Well, there's the smoking gun ... somebody marked (27,2) as
XMIN_COMMITTED, showing that they thought 46034931 was committed, while
someone else marked (27,43) as XMAX_INVALID, showing that they thought
46034931 was aborted. So we have some kind of very-infrequent
breakage in transaction commit-state lookup. Or a hardware problem,
but I suspect we are looking at a bug.
Could you check out what pg_clog has for transaction 46034931?
This would be pg_clog/002B (which dates your problem to Dec 29 BTW),
byte at offset 39BFC hex or 236540 decimal. I forget which way the
bits run within the byte but will look it up if you can get me the
value of that byte.
I'm off to take a real close look at what was done to the pg_clog code
during the 7.4 cycle ...
regards, tom lane
