No brtfs. We are going to try turning off synchronous_commit temporarily to see if there are underlying I/O issues. On Mon, 24 May 2021 at 22:59, Alexey M Boltenkov <padrebolt@xxxxxxxxx> wrote: > > On 05/24/21 19:24, Christophe Pettus wrote: > > > >> On May 24, 2021, at 09:22, Bob Jolliffe <bobjolliffe@xxxxxxxxx> wrote: > >> > >> It is hard to say as it only happens for 30s couple of times per day. > >> Everything does return to normal after the blocking transaction is > >> committed. It could be a disk thing or even a network issue (the java > >> app is on a different machine to the db). But I never saw > >> transactions blocked in commit before so was wondering if there is any > >> rational set of reasons why it might do that. > > One thing you can check is to turn off synchronous_commit (understanding the possibility of "time loss" in the event of a system crash). If that mitigates the problem, the issue is likely the I/O subsystem blocking during the fsync() operation. > > > > > Just a question. Is there a btrfs(with compression maybe) around? 30 > seconds is a commit(file system) timeout for btrfs. Some processes like > btrfs cleaner/allocate/worker on top of CPU/io use? >
