Re: basebackup checksum verification

Greetings,

* Tomas Vondra (tomas(dot)vondra(at)2ndquadrant(dot)com) wrote:
> On Tue, Mar 26, 2019 at 04:49:21PM -0700, Andres Freund wrote:
> >On 2019-03-26 19:22:03 -0400, Stephen Frost wrote:
> >>* Andres Freund (andres(at)anarazel(dot)de) wrote:
> >>> As detailed in
> >>> https://postgr.es/m/20190319200050.ncuxejradurjakdc%40alap3.anarazel.de
> >>> the way the backend's basebackup checksum verification works makes its
> >>> error detection capabilities very dubious.
> >>
> >>I disagree that it's 'very dubious', even with your analysis.
> >
> >I really don't know what to say. The current algorithm is flat out
> >bogus.
>
> Bogus might be a bit too harsh, but yeah - failure to reliably detect obviously
> invalid checksums when the LSN just happens to be high due to randomness is not
> a good thing. We'll still detect pages corrupted in other places, but this is
> rather unfortunate.

I'm all for improving it, as I said originally.

> >>I thought Robert's response was generally good, pointing out that
> >>we're talking about this being an issue if the corruption happens in a
> >>certain set of bytes. That said, I'm happy to see improvements in
> >>this area but I'm flat out upset about the notion that we must be
> >>perfect here- our checksums themselves aren't perfect for catching
> >>corruption either.
> >
> >The point is that we're not detecting errors that we can detect when
> >read outside of basebackup. I really entirely completely fail how that
> >can be defended.
> >
> >I think we're making promises with this the basebackup feature we're not
> >even remotely keeping. I don't understand how you can defend that, given
> >the current state, you can have a basebackup that you took with
> >checksums enabled, and then when actually use that basebackup you get
> >checksum failures. Like it's one thing not to detect all storage
> >issues, but if we do detect them after using the basebackup, that's
> >really not ok.
>
> Yeah, if basebackup completes without reporting any invalid checksums, but
> running pg_verify_checksums on the same backups detects those, that probably
> should raise some eyebrows.

That isn't actually what would happen at this point, just so we're
clear. What Andres is talking about is a solution which would only
actually work for pg_basebackup, and not for pg_verify_checksums
(without some serious changes which make it connect to the running
server and run various functions to perform the locking that he's
proposing pg_basebackup do...).

> We already have such blind spot, but it's expected to be pretty small
> (essentially pages modified since start of the backup).

eh..? This is.. more-or-less entirely what's being discussed here:
exactly how we detect and determine which pages were modified since the
start of the backup, and which might have been partially written out
when we tried to read them and therefore fail a checksum check, but it
doesn't matter because we don't actually end up using those pages.

I outlined a couple of other approaches to improving that situation,
which would be able to be used with pg_verify_checksums without having
to connect to the backend, but I'll note that those were completely
ignored, leading me to believe that there's really not much more to
discuss here since other ideas are just not open to being considered.

Thanks,

Stephen