Re: stat() vs ERROR_DELETE_PENDING, round N + 1

Thomas Munro <thomas(dot)munro(at)gmail(dot)com> writes:
> On Thu, Sep 2, 2021 at 10:31 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> That seems quite horrid :-(. But if it works, doesn't that mean that
>> somewhere we are opening a problematic file without the correct
>> sharing flags?

> I'm no expert, but not AFAICS. We managed to delete the file while
> some other backend had it open, which FILE_SHARE_DELETE allowed. We
> just can't open it or create a new file with the same name until it's
> really gone (all handles closed).

Right, but we shouldn't ever need to access such a file --- we couldn't do
so on Unix, certainly. So for the open() case, it's sufficient to return
ENOENT, and the problem is only to make sure that that's what we return if
the underlying error is ERROR_DELETE_PENDING.

It's harder if the desire is to create a new file of the same name.
I'm inclined to think that the best answer might be "if it hurts,
don't do that". We should not have such a case for ordinary relation
files or WAL files, but maybe there are individual other cases where
some redesign is indicated?

regards, tom lane