> 20 июля 2020 г., в 21:44, Alvaro Herrera <alvhe...@2ndquadrant.com>
> написал(а):
>
>> I think we shall do that in some cases
>> but IMHO it's not a very good idea in all the cases. Basically, if
>> the xmin precedes the relfrozenxid then probably we should allow to
>> update the relfrozenxid whereas if the xmin precedes cutoff xid and
>> still uncommitted then probably we might stop relfrozenxid from being
>> updated so that we can stop CLOG from getting truncated.
>
> I'm not sure I understand 100% what you're talking about here (the first
> half seems dangerous unless you misspoke), but in any case it seems a
> pointless optimization. I mean, if the heap is corrupted, you can hope
> to complete the vacuum (which will hopefully return which *other* tuples
> are similarly corrupt) but trying to advance relfrozenxid is a lost
> cause.
I think the point here is to actually move relfrozenxid back. But the mince
can't be turned back. If CLOG is rotated - the table is corrupted beyond easy
repair.
I'm not sure it's Dilip's case, but I'll try to describe what I was
encountering.
We were observing this kind of corruption in three cases:
1. With a bug in patched Linux kernel page cache we could loose FS page write
2. With a bug in WAL-G block-level incremental backup - we could loose update
of the page.
3. With a firmware bug in SSD drives from one vendor - one write to block
storage device was lost
One page in a database is of some non-latest version (but with correct
checksum, it's just an old version). And in our case usually a VACUUMing of a
page was lost (with freezes of all tuples). Some tuples are not marked as
frozen, while VM has frozen bit for page. Everything works just fine until
someone updates a tuple on the same page: VM bit is reset and eventually user
will try to consult CLOG, which is already truncated.
This is why we may need to defer CLOG truncation or even move relfrozenxid back.
FWIW we coped with this by actively monitoring this kind of corruption with
this amcheck patch [0]. One can observe this lost page updates cheaply in
indexes and act on first sight of corruption: identify source of the buggy
behaviour.
Dilip, does this sound like a corruption case you are working on?
Thanks!
Best regards, Andrey Borodin.
[0] https://commitfest.postgresql.org/24/2254/
