This commit makes me very uneasy.  I didn't much care about this patch
mainly because I didn't imagine its consequences. Now, I see following:

1) We uglify buffer manager interface a lot.  This patch adds 3 more
arguments to BufferGetPage().  It looks weird.  Should we try to find less
invasive way for doing this?
2) Did you ever try to examine performance regression?  I tried simple read
only test on 4x18 Intel server.
pgbench -s 1000 -c 140 -j 100 -M prepared -S -T 300 -P 1 postgres (data
fits to shared_buffers)

master                    -   193 740.3 TPS
snapshot too old reverted - 1 065 732.6 TPS

So, for read-only benchmark this patch introduces more than 5 times
regression on big machine.

Alexander Korotkov
Postgres Professional:
The Russian Postgres Company

Reply via email to