On Sat, Feb 4, 2017 at 4:33 AM, Andres Freund <andres@anarazel.de> wrote:
On 2017-02-03 19:13:45 -0600, Jim Nasby wrote: > No, I noticed it while reading code. Removing that does mean that if any > non-default strategy (in any backend) hits that buffer again then the buffer > will almost certainly migrate into the main buffer pool the next time one of > the rings hits that buffer
Well, as long as the buffer is used from the ring, BufferAlloc() - BufferAlloc() will reset the usagecount when rechristening the buffer. So unless anything happens inbetween the buffer being remapped last and remapped next, it'll be reused. Right?
The only case where I can see the old logic mattering positively is for synchronized seqscans. For pretty much else that logic seems worse, because it essentially prevents any buffers ever staying in s_b when only ringbuffer accesses are performed.
I'm tempted to put the old logic back, but more because this likely was unintentional, not because I think it's clearly better.
+1
Yes, it was unintentional change. So we should put old logic back unless we have proof that this change make it better.
Patch is attached. I tried to make some comments, but probably they are not enough.