On 10/27/2012 2:41 PM, Heikki Linnakangas wrote:
> And it's not at all
> clear to me that it would perform better than full_page_writes. You're
> writing and flushing out roughly the same amount of data AFAICS.
I think this assumption is wrong. full_page_writes=on means we write the
full page content to WAL on first change after a checkpoint. A change
after a checkpoint logically means that the same page is dirty now and
must also be written latest during the next checkpoint, which means 16K
written minimum for every page changed after a checkpoint.
> What exactly is the problem with full_page_writes that we're trying to
> solve?
Full page writes are meant to guard against torn pages. That is the case
when an 8K page is written by the underlying OS/filesystem/HW in smaller
chunks (for example 512 byte sectors), and in the case of a crash some
of these chunks make it, others don't. Without full_page_writes, crash
recovery can work if all 8K made it, or nothing made it (aka atomic
writes). But it will fail otherwise.
The amount of WAL generated with full_page_writes=on is quite
substantial. For pgbench for example the ratio 20:1. Meaning with
full_page_writes you write 20x the amount you do without.
Jan
--
Anyone who trades liberty for security deserves neither
liberty nor security. -- Benjamin Franklin