Hi,
On 2022-01-24 09:42:13 +1300, Thomas Munro wrote:
> On Sun, Jan 23, 2022 at 7:52 AM Noah Misch <noah@leadboat.com> wrote:
> > Future work can benchmark the new behavior and, if it performs well, make
> > it unconditional in v15+. I would expect performance to be unchanged or
> > slightly better, because the new behavior requests less futile work from
> > the OS.
I doubt it'll be generally applicable. Turning a write operation into a
read-write isn't free. Yes, often enough it's likely that the prior page is
still in cache, but I don't think we can rely on that in general.
It also just fundamentally locks us into never using O_DIRECT in anger. I
don't think that's a good direction.
> One detail is that wal_level=open_datasync, wal_senders=0,
> wal_level=minimal will panic, because O_DIRECT requires fs
> page-aligned access (and fails in various other ways on other OSes, eg
> expensive read-before-write every time). That's an ultra-niche
> concern likely affecting nobody, especially when multiplied by the
> odds that anyone is using that stack at all (considering that
> streaming rep has apparently been borked for years on linux/sparc/ext4
> and nobody told us).
Seems like the patch should at least make this error out?
Greetings,
Andres Freund