Re: Checking for missing heap/index files

On Thu, Jun 9, 2022 at 09:46:51AM -0700, Mark Dilger wrote:
>
>
> > On Jun 8, 2022, at 5:45 AM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> >
> > Is this something anyone has even needed or had requested?
>
> I might have put this in amcheck's verify_heapam() had there been an
> interface for it. I vaguely recall wanting something like this, yes.
>
> As it stands, verify_heapam() may trigger mdread()'s "could not open
> file" or "could not read block" error, in the course of verifying
> the table. There isn't an option in amcheck to just verify that
> the underlying files exist. If struct f_smgr had a function for
> validating that all segment files exist, I may have added an option to
> amcheck (and the pg_amcheck frontend tool) to quickly look for missing
> files.

Well, how do we know what files should exist? We know the first segment
should probably exist, but what about +1GB segments?

> Looking at smgr/md.c, it seems mdnblocks() is close to what we want,
> but it skips already opened segments "to avoid redundant seeks".
> Perhaps we'd want to add a function to f_smgr, say "smgr_allexist",
> to check for all segment files? I'm not sure how heavy-handed the
> corresponding mdallexist() function should be. Should it close
> all open segments, then reopen and check the size of all of them
> by calling mdnblocks()? That seems safer than merely asking the
> filesystem if the file exists without verifying that it can be opened.

Yes.

> If we made these changes, and added corresponding quick check options
> to amcheck and pg_amcheck, would that meet your current needs? The
> downside to using amcheck for this sort of thing is that we did not
> (and likely will not) back port it. I have had several occasions to
> want this functionality recently, but the customers were on pre-v14
> servers, so these tools were not available anyway.

I don't have a need for it --- I was just wondering why we have
something that checks the relation contents, but not the file existence?

It seems like pg_amcheck would be a nice place to add this checking
ability.

--
Bruce Momjian <bruce(at)momjian(dot)us> https://momjian.us
EDB https://enterprisedb.com

Indecision is a decision. Inaction is an action. Mark Batterson