query-source port 53;

  zone "linux.bogus" {
        type slave;
        file "sz/linux.bogus";
        masters { 127.0.0.1; };
  };

There are four items regarding this:

I have discovered with newer versions of BIND that this
[<em/shuffeling files, -ed/] is no longer necessary.  There is a
"forward" directive in addition to the "forwarders" directive that
controls how they are used.  The default setting is "forward first",
which first asks each of the forwarders, and then tries the normal
approach of doing the legwork itself if that fails.  This gives the
familiar behavior of gethostbyname() taking an inordinately long time
when the link is not up.  But if "forward only" is set, then BIND
gives up when it doesn't get a response from the forwarders, and
gethostbyname() returns immediately.  Hence there is no need to
perform sleight-of-hand with files in /etc and restart the server.

In my case, I just added the lines

forward only;
forwarders { 193.133.58.5; };

to the options { } section of my named.conf file. It works very
nicely. The only disadvantage of this is that it reduces an incredibly
sophisticated piece of DNS software to the status of a dumb cache. To
some extent, I would just like to run a dumb cache for DNS instead,
but there doesn't seem to be such a piece of software available for
Linux.

I run named on my 'Masquerading' machine here. I have 
two root.hints files, one called root.hints.real which contains 
the real root server names and the other called root.hints.fake 
which contains...

----
; root.hints.fake
; this file contains no information
----

When I go off line I copy the root.hints.fake file to root.hints and
restart named.

When I go online I copy root.hints.real to root.hints and restart
named.

This is done from ip-down & ip-up respectively.

The first time I do a query off line on a domain name named doesn't
have details for it puts an entry like this in messages..

Jan 28 20:10:11 hazchem named[10147]: No root nameserver for class IN

which I can live with.
         
It certainly seems to work for me. I can use the nameserver for
local machines while off the 'net without the timeout delay for
external domain names and I while on the 'net queries for external
domains work normally

Peter Denison thought that Ian does not go far enough though. He writes:

When connected) serve all cached (and local network) entries immediately
                for non-cached entries, forward to my ISPs nameserver
When off-line)  serve local network queries immediately
                fail all other queries **immediately**

The combination of changing the root cache file and forwarding queries
doesn't work.

So, I've set up (with some discussion of this on the local LUG) two nameds
as follows:

named-online:   forwards to ISPs nameserver
                master for localnet zone
                master for localnet reverse zone (1.168.192.in-addr.arpa)
                master for 0.0.127.in-addr.arpa
                listens on port 60053

named-offline:  no forwarding
                "fake" root cache file
                slave for 3 local zones (master is 127.0.0.1:60053)
                listens on port 61053

And combined this with port forwarding, to send port 53 to 61053 when
off-line, and to port 60053 when online. (I'm using the new netfilter
package under 2.3.18, but the old (ipchains) mechanism should work.)

Note that this won't quite work out-of-the-box, as there's a slight bug in
BIND 8.2, which I have logged wth the developers, preventing a slave
having a master on the same IP address (even if a different port). It's a
trivial patch, and should go in soon I hope.

I use to run my own named on all my machines which are only
occasionally connected to the Internet by modem. The nameserver only
acts as a cache, it has no area of authority and asks back for
everything at the name servers in the root.cache file. As is usual
with Slackware, it is started before nfsd and mountd.

With one of my machines (a Libretto 30 notebook) I had the problem
that sometimes I could mount it from another system connected to my
local LAN, but most of the time it didn't work.  I had the same effect
regardless of using PLIP, a PCMCIA ethernet card or PPP over a serial
interface.

After some time of guessing and experimenting I found out that
apparently named messed with the process of registration nfsd and
mountd have to carry out with the portmapper upon startup (I start
these daemons at boot time as usual). Starting named after nfsd and
mountd eliminated this problem completely.

As there are no disadvantages to expect from such a modified boot
sequence I'd advise everybody to do it that way to prevent potential
trouble.

	Specific to BIND 8/9, Adam L Rice has sent me this e-mail, about how to run DNS painlessly on a dialup machine:
	I have received this mail from Ian Clark <ic@deakin.edu.au> where he explains his way of doing this:
	I have also received information about how BIND interacts with NFS and the portmapper on a mostly offline machine from Karl-Max Wanger: