"Hiroshi Inoue" <Inoue(at)tpf(dot)co(dot)jp> writes:
> AFAIR t_ctid isn't logged in WAL.
After looking at the heap_update code I think you are right. Doesn't
that render the field completely useless/unreliable?
In the simple heap_update case I think that heap_xlog_update could
easily set the old tuple's t_ctid field correctly. Not sure how
it works when VACUUM is moving tuple chains around, however.
Another thing I am currently looking at is that I do not believe VACUUM
handles tuple chain moves correctly. It only enters the chain-moving
logic if it finds a tuple that is in the *middle* of an update chain,
ie, both the prior and next tuples still exist. In the case of a
two-element update chain (only the latest and next-to-latest tuples of
a row survive VACUUM), AFAICT vacuum will happily move the latest tuple
without ever updating the previous tuple's t_ctid.
In short t_ctid seems extremely unreliable. I have been trying to work
out a way that a bad t_ctid link could lead to the duplicate-tuple
reports we've been hearing lately, but so far I haven't seen one. I do
think it can lead to missed UPDATEs in read-committed mode, however.
regards, tom lane
In response to
pgsql-hackers by date
|Next:||From: Peter Eisentraut||Date: 2001-12-22 21:41:51|
|Subject: HISTORY file|
|Previous:||From: Hiroshi Inoue||Date: 2001-12-22 15:52:37|
|Subject: Re: Possible bug in vacuum redo|