Re: Load distributed checkpoint

Gregory Stark <stark(at)enterprisedb(dot)com> writes:
> It's a fundamental shift in the idea of the purpose of bgwriter. Instead of
> trying to suck i/o away from the subsequent checkpoint it would be responsible
> for all the i/o of the previous checkpoint which would still be in progress
> for the entire time of checkpoint_timeout. It would only complete when
> bgwriter had finished doing its one full sweep.

I think that's basically the same as the original suggestion, which is
to do checkpoints using less than the full I/O bandwidth of the machine
--- tying it exactly to the default bgwriter rate may or may not be
appropriate.

The difficulty with such schemes is that if you go over to using O_DIRECT
writes instead of fsync in the bgwriter, it's hard to avoid doing the
same when a random backend has to write a dirty buffer --- yet you'd
really rather that such a backend not have to wait for the ensuing I/O.
And this gets a lot worse if checkpoints are slowed down, because it gets
more likely that the bufmgr will run out of clean buffers and have to do
a write() from a backend.

regards, tom lane