Re: Discouraging :local:

[Greg A. Woods]

[ On Friday, January 24, 2003 at 21:57:57 (-0800), Kenneth Porter wrote: ]
> Subject: Re: Discouraging :local:
>
> --On Thursday, January 23, 2003 12:43 PM -0500 Larry Jones
> <address@hidden> wrote:
> 
> > What makes :local: inadvisable is the disk not being local, but rather
> > being on some kind of network filesystem.  I don't know of any way to
> > detect that.
> 
> But *why* is that bad? After all, a SCSI disk is on the other end of a SCSI
> cable, and so is "networked" in some sense. Why is that ok but a "network"
> disk is not?

A disk on the end of a SCSI cable, or even fibre-channel SAN, really is
"local" in terms of the filesystem and the way the host treats the
device.

A network disk is not just not directly attached to the host in
question, it's also using a different kind of access protocol and
filesystem, and often it may also be sharing the remote files with other
hosts.

The latter, the sharing part, is where the real trouble begins.
Ensuring reliable order of operations for various operations which would
be "atomic" on a local filesystem is very very difficult (literally
impossible in some cases) for shared network filesystems.

Also of course there are potential data integrity issues with networked
filesystems.  The integrity checks in UDP are really quite lame.  The
integrity checks in TCP are not even all that great.  Meanwhile SCSI
parity checks are more reliable and the nature of how hardware errors
can occur in something like a SCSI interface make those parity checks
more likely to detect all errors.

-- 
                                                                Greg A. Woods

+1 416 218-0098;            <address@hidden>;           <address@hidden>
Planix, Inc. <address@hidden>; VE3TCP; Secrets of the Weird <address@hidden>