So Debian bug #710747 led to an interesting discussion the other night on IRC which made me realize there are a lot of people who have yet to understand why sysvinit needs replacing.

I can't speak to whether upstart solves this bug in particular. The tftpd-hpa package in Ubuntu (and in Debian experimental) does have an upstart job, and I have used tftpd-hpa on systems whose network interfaces are managed entirely with NetworkManager, and I have not seen this bug; but I can't say that this is more than coincidence. However, I can speak to how upstart provides facilities that address the general problem of starting services that require a working network connection to be present before they start up.

Traditionally on Linux systems, init scripts could assume that the network connection is up before starting any services for the target runlevel. With LSB dependencies, you could also express this as

Required-Start:    $network

The trouble with both of these approaches is that they assume that there is a single point in time, early in the boot of the system, when the "network" is up. This was an ok assumption 15 years ago, when the Linux kernel initialized all devices serially and all our network configuration was static. But nowadays, you simply cannot rely on systems booting in such a way that the network is fully up shortly after boot - or in a way that you can block the system boot waiting for the network to be up.

If all of our services would cope gracefully with being started without the network, this would be a non-issue. Unfortunately, not all network services are written in such a way as to work without a network; nor do they all cope with dynamic changes to interfaces or addresses. While it would be better if these services were written more robustly, since there's no shortage of daemons written this way, it's convenient that upstart provides tools for coping with this behavior in the meantime.

Here's a real (simplified) example of an upstart job for a service that needs to wait for a non-loopback interface before it starts:

start on (local-filesystems and net-device-up IFACE!=lo)
stop on runlevel [!2345]

expect fork
exec nmbd -D

Now, you might also have a service that you only want to start up when a particular network device comes up. This is easily expressed in upstart, and might look like this hypothetical job:

start on net-device-up IFACE=wlan0
stop on net-device-down IFACE=wlan0

expect daemon
exec mydaemon

If you need to restart a daemon whenever the set of active network connections changes, that's less straightforward. Upstart doesn't have the notion of a 'restart' rule, so you would need two jobs to handle this - one for the daemon itself that starts on the first network connection, and a second job to trigger a restart of the daemon when the network status changes. For the tftpd-hpa case in the abovementioned bug, this might look like:

$ cat /etc/init/tftpd-hpa.conf
start on runlevel [2345]
stop on runlevel [!2345]

expect fork

        if [ -f ${DEFAULTS} ]; then
                . ${DEFAULTS}
        exec /usr/sbin/in.tftpd --listen  --user ${TFTP_USERNAME} --address ${TFTP_ADDRESS} ${TFTP_OPTIONS} ${TFTP_DIRECTORY}
end script
$ cat /etc/init/tftpd-hpa-restart.conf
start on net-device-up
instance $IFACE

        status tftpd-hpa | grep -q start/ || stop
        restart tftpd-hpa
end script

For this case, upstart doesn't provide quite so great an advantage. It's nice to be able to use upstart natively for both pieces, but you can do the same thing with an init script plus an if-up.d script for ifupdown, which is what maintainers do today. I think adding a restart on stanza to upstart in the future to address this would be a good idea. Though in any event, this is far simpler to do with upstart than with any if-up.d script I've ever seen for managing an initscript-based service. Between the more friendly declarative syntax, and the enhanced semantics for expressing when to run (or restart) a job, upstart offers clear advantages over other init systems.