Re: Checkpoint cost, looks like it is WAL/CRC

Zeugswetter Andreas DAZ SD wrote:
>
> >> Are you sure about that? That would probably be the normal case, but
> >> are you promised that the hardware will write all of the sectors of a
>
> >> block in order?
> >
> > I don't think you can possibly assume that. If the block
> > crosses a cylinder boundary then it's certainly an unsafe
> > assumption, and even within a cylinder (no seek required) I'm
> > pretty sure that disk drives have understood "write the next
> > sector that passes under the heads"
> > for decades.
>
> A lot of hardware exists, that guards against partial writes
> of single IO requests (a persistent write cache for a HP raid
> controller for intel servers costs ~500$ extra).
>
> But, the OS usually has 4k (some 8k) filesystem buffer size,
> and since we do not use direct io for datafiles, the OS might decide
> to schedule two 4k writes differently for one 8k page.
>
> If you do not build pg to match your fs buffer size you cannot
> guard against partial writes with hardware :-(
>
> We could alleviate that problem with direct io for datafiles.

Now that is an interesting analysis. I thought people who used
batter-backed drive cache wouldn't have partial page write problems, but
I now see it is certainly possible.

--
Bruce Momjian | http://candle.pha.pa.us
pgman(at)candle(dot)pha(dot)pa(dot)us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073