Re: Experimental ARC implementation

Jan Wieck <JanWieck@Yahoo.com> writes:
> I am not really aiming at removing sync() alltogether.
> ...
> What doing frequent fdatasync/fsync during a constant ongoing checkpoint 
> will cause is to significantly lower the physical write storm happening 
> at the sync(), which is causing huge problems right now.

That's fair as far as solving the performance issue is concerned.  But
I think if we are going to muck about in this area, we ought to see
whether we can't remove the use of sync() altogether, because of the
issue of reliability.  When you depend on sync() for a checkpoint, you
really can't be sure that you are preserving the fundamental WAL
invariant --- you don't *know* that all the data writes have been done
before you write the checkpoint record to WAL.

I realize that reducing the amount of I/O that sync() triggers
would make this a lot less of a risk in practice than it is now, but
I'd still be happier if we could eliminate the risk rather than just
reduce it.
        regards, tom lane