Marko Tiikkaja <marko@joh.to> writes:
> On 1/8/14 4:55 PM, Tom Lane wrote:
>> If I did the math right, this corresponds to pg_clog file 010C. Do you
>> have such a file? Does it look to have the right size and permissions?
>> What file names are present in pg_clog?
> No, it's not there. The files look like this:
> -rw------- 1 postgres postgres 256K Nov 3 15:55 0390
> -rw------- 1 postgres postgres 256K Nov 3 22:26 0391
> ..
> -rw------- 1 postgres postgres 256K Jan 8 16:43 044D
> -rw------- 1 postgres postgres 24K Jan 8 17:28 044E
Oh, I fat-fingered the math somehow, because when I do it over again,
I get 0430 (or 0438 for your other observed XID). That's right in the
range of your available clog files, so the notion of a crazy XID in
the message queue doesn't seem to be the explanation here.
Since the referenced XIDs are not close together, I assume that the two
traces also show different pagenos (SimpleLruReadPage) and slotnos
(SimpleLruWaitIO) and lockids (LWLockAcquire)?
That leads to the idea that we've got some condition that broke
all of the clog SLRU slots at the same time, though it's really hard
to see what that might be ...
regards, tom lane