On Tue, Mar 12, 2019 at 10:49 PM Nicola Contu <nicola.contu@xxxxxxxxx> wrote: > Il giorno gio 7 mar 2019 alle ore 09:39 Nicola Contu <nicola.contu@xxxxxxxxx> ha scritto: >> So the first file is on Postgres11.2 on a test server (and where I compare 10 vs 11) >> The second file, is our preprod machine running Postgres 11.2 (different hardware etc, it is a VM). I know that could be confusing, but I just wanted to compare that too because if you see the two files there's a lot of difference between the two machines. >> And they are both running CentOS 7. >> >> So at this point I have two problems. One inside the machine between Postgres 10 and 11 and another problem on the preprod (similar to prod) with a lot of lseek. So we still don't know what is different between 10 and 11. We'll need some more evidence to understand that. We know that perf said 11 was spending more time making syscalls, but we haven't seen any syscall measurement from the 10 system so we don't know what those supposed extra syscalls are. Also, backing up a bit, have you compared the query plan to see if it's the same, and can we please see it? As for the preprod/lots-of-lseek system, I don't know, it seems there are even more unknown variables there... if you can say more about the virtualisation technology you're using, perhaps someone who knows more about that would have some ideas. Besides apparently slower syscalls, one factor that is unexplained is why that system is calling lseek() more times per query (but you said there may be other work happening on the server, so who knows). -- Thomas Munro https://enterprisedb.com
