(2) Go back to timestamp as certificate revision number. This is simpler
and more robust than using the network revision number for this and
forcing network revision fast-forward, which could cause some peers
to fall off the horizon when you don't want them to.
and go ahead and bump version to 1.0.4.
For a while in 1.0.3 -dev I was trying to optimize out repeated network controller
requests by using a ratcheting mechanism. If the client received a network config
that was indeed different from the one it had, it would respond by instantlly
requesting it again.
Not sure what I was thinking. It's fundamentally unsafe to respond to a message
with another message of the same type -- it risks a race condition. In this case
that's exactly what could happen.
It just isn't worth the added complexity to avoid a tiny, tiny amount of network
overhead, so I've taken this whole path out.
A few extra bytes every two minutes isn't worth fretting about, but as I recall
the reason for this optimization was to save CPU on the controller. This can be
achieved by just caching responses in memory *there* and serving those same
responses back out if they haven't changed.
I think I developed that 'ratcheting' stuff before I went full time on this. It's
hard to develop stuff like this without hours of sustained focus.
Adam Ierymenko
Kees Bos