Re: NetBSD on a database server - WAPBL, softdeps

["Steven M. Bellovin" <smb%cs.columbia.edu@localhost>]

On Fri, 18 Sep 2009 01:21:06 +0100
Mindaugas Rasiukevicius <rmind%NetBSD.org@localhost> wrote:

> Thor Lancelot Simon <tls%panix.com@localhost> wrote:
> > >
> > > and keeping in mind softdeps are scheduled to be removed in 6.0  
> > > (http://mail-index.netbsd.org/netbsd-announce/2008/12/14/msg000051.html), 
> > > what is the recommended configuration, filesystem wise, for a
> > > database server running on NetBSD? Wouldn't running with "bare"
> > > UFS be too slow?
> > 
> > Why would you expect any benefit from either WAPBL or soft updates
> > on a database server?  Databases (or, at least, any database worth
> > using) manage their tablespaces independently, generally without
> > creating or deleting files in the filesystem -- or doing so very
> > seldom, if at all.
> 
> Are you suggesting that long boots after system failure is the option?
> I think journaling for larger databases is mandatory.  And advisable
> for any other case.

The issue Thor is pointing out is the granularity of recovery.  WAPBL
and fsck use file-level granularity -- they ensure the consistency of
the *file system*.  Databases have their own, complex internal
structure; WAPBL and fsck know nothing of them.  If you want
consistency, you have to ensure that the important blocks are written
out in the proper order.  It's reasonable to want the database to do
that; it is not reasonable to expect the system to.

I think the system call you should be concerned about is
fsync_range().  Is that used (properly) by your database?  Is it fast
enough?
> 
> If there are issues with our fsync(), we ought to fix them.  And
> perhaps invest some time in optimisations, like Oracle invested time
> to decrease fsync() latency on ext3, recently.
> 
See above.


                --Steve Bellovin, http://www.cs.columbia.edu/~smb