Picking back up where this left off last year, because I apparently only
work on TACACS during the holidays :)
On 12/30/2013 7:28 PM, Jimmy Hess wrote:
Even 5 seconds extra for each command may hinder operators, to the extent
it would be intolerable; shell commands should run almost
instantaneously.... this is not a GUI, with an hourglass. Real-time
responsiveness in a shell is crucial --- which remote auth should not
change. Sometimes operators paste a buffer with a fair number of
commands, not expecting a second delay between each command --- a
repeated delay, may also break a pasted sequence.
It is very possible for two of three auth servers to be unreachable, in
case of a network break, but that isn't necessary. The "response
timeout" might be 5 seconds, but in reality, there are cases where you
would wait longer, and that is tragic, since there are some obvious
alternative approaches that would have had results that would be more
'friendly' to the interactive user.
(Like remembering which server is working for a while, or remembering
that all servers are down -- for a while, and having a 50ms timeout,
with all servers queried in parallel, instead of a 5 seconds timeout)
I think this needs to be part of the specification.
I'm sure the reason they didn't do parallel queries was because of both
network and CPU load back when the protocol was drafted. But it might
be good to have local caching of authentication so that can happen even
when servers are down or slow. Authorization could be updated to send
the permissions to the router for local handling. Then if the server
dies while a session is open only accounting would be affected.
That does increase the vendors/implementors work but it might be doable
in phases and with partial support with the clients and servers
negotiating what is possible. The biggest drawback to making things
like this better is you don't gain much except during outages and if you
increase complexity too much you make it wide open for bugs.
Maybe there is a simpler solution that keeps you happy about redundancy
but doesn't increase complexity that much (possibly anycast tacacs, but
the session basis of the protocol has always made that not feasible).
It's possible that one of the L4 protocols Saku Ytti mentioned, QUIC or
MinimaLT would address these problems too. It's possible that if we did
the transport with BEEP it would also provide this, but I'm reading the
docs and I don't think it goes that far in terms of connection assurance.
--
-JH
So, here is my TACACS RFC christmas list:
1. underlying crypto
2. ssh host key authentication - having the router ask tacacs for an
authorized_keys list for rdrake. I'm willing to let this go because
many vendors are finding ways to do key distribution, but I'd still like
to have a standard (https://code.google.com/p/openssh-lpk/ for how to do
this over LDAP in UNIX)
3. authentication and authorization caching and/or something else
