Hi,
On 2022-05-28 13:13:20 +1200, Thomas Munro wrote:
> There was a little bit of discussion on #pgcon-stream2 which I could
> summarise as: can we figure out a way to keep parts of the CLOG pinned
> so that backends don't have to do that for each lookup? Then CLOG
> checks become simple reads.
Included in that is not needing to re-check that the identity of the buffer
changed since the last use and to not need a PrivateRefCountEntry. Neither is
cheap...
I'd structure it so that there's a small list of slru buffers that's pinned in
a "shared" mode. Entering the buffer into that increases the BufferDesc's
refcount, but is *not* memorialized in the backend's refcount structures,
because it's "owned by the SLRU".
> There may be some relation to the idea of
> 'nailing' btree root pages that I've heard of from a couple of people
> now (with ProcSignalBarrier or something more fine grained along those
> lines if you need to unnail anything). Something to think about.
I'm very doubtful it's a good idea to combine those things - I think it's
quite different to come up with a design for SLRUs, of which there's a
constant number and shared memory ownership datastructures, and btree root
pages etc, of which there are arbitrary many.
For the nbtree (and similar) cases, I think it'd make sense to give backends a
size-limited number of pages they can keep pinned, but in a backend local
way. With, as you suggest, a procsignal barrier or such to force release.
> I'm also wondering if it would be possible to do "optimistic" pinning
> instead for reads that normally need only a pin, using some kind of
> counter scheme with read barriers to tell you if the page might have
> been evicted after you read the data...
-many
That seems fragile and complicated, without, at least to me, a clear need.
Greetings,
Andres Freund