[Buildbot-devel] Slaves behind Dynamic IPs

[Brian Warner]

> I have a couple of slaves behind dynamic IP addresses.  When the IP
> address changes they are no longer able to talk to the master, getting
> this traceback:

Interesting. I've seen some related problems on the Twisted buildbot, where a
slow link and a large logfile cause the slave to disconnect+reconnect in a
way that makes the buildmaster think there are two slaves vying for the same
builder: not unlink the address-changed problem.

I'm currently rewriting the lost-slave detection code, starting with some
test cases to handle these situations. Hopefully that will be able to
accomodate this problem.

Do you have any idea how frequently the address is changing? Is it a question
of your NAT box timing out an active connection and then giving the
reconnection-attempt a different source address? NAT boxes really shouldn't
give up on active connections. There's a buildslave-side option to basically
ping the master every N seconds, which might help avoid a NAT timeout.. try
--keepalive=300 to ping the master every 5 minutes.

If you've got some kind of dialup connection that gets spontaneously
disconnected every once in a while.. hmm. Well, maybe this overhaul will
help. In the meantime the 'ping builder' button might accelerate the process
of discovering that the slave has gone away, and once the master gives up on
it then the incoming slave is free to claim the builder that the old one left
behind.

thanks,
 -Brian