On Wed, May 11, 2022 at 05:23:27PM +0200, Alvaro Herrera wrote:
> As I understand it, the point of using a ring is to throttle performance
> for bulk operations such as vacuum. I'm not sure why we would want to
> throttle either MERGE or INSERT; it seems to me that we should want them
> to go as fast as possible.
The point of the ring/strategy buffer is to avoid a seq scan/vacuum/COPY
clobbering the entire buffer cache when processing a table larger than
shared_buffers.
It also makes backends doing bulk operations responsible for their own writes,
rather than leaving other backends to clean up all their dirty buffers.
> If MERGE were to use a ring buffer, why wouldn't UPDATE do the same?
> There are some comments to that effect in src/backend/buffer/README --
> they do mention UPDATE/DELETE and not INSERT. It seems to me that these
> three commands (MERGE/UPDATE/DELETE) should be handled in similar ways,
> so I don't think we need to consider lack of MERGE using a ring buffer
> an open item for pg15.
I tend to think that a bulk commad either should or, at least, should be *able*
to use a ring buffer. The README doesn't consider INSERTs, but that's the one
that we'd be most interested in. I guess you're right that MERGE ought to do
what the existing commands are doing.
--
Justin