Hi Andres,
> No fsyncs will be issued in the common paths, and they won't be written
> to disk by !shutdown checkpoints. But they *will* be written to disk if
> there's not enough space in shared_buffers for all the data.
With unlogged tables I still see data written to the device every 10
seconds, so I did some digging and found bgwriter_delay = 10000, again
with an artifical 10s limit applied on the parameter. To disable
background writing, I set bgwriter_lru_maxpages = 0 which seems to
help.
What is left, is the wal-writer-process which still seems to fsync
every 10s, even though there is no data written to any logged table.
Regards, Clemens