Re: [squid-users] Too many TCP_DENIED/407 when using Kerberos authentication

From: Brendan Kearney <bpk678_at_gmail.com>
Date: Tue, 01 Oct 2013 07:47:52 -0400

On Tue, 2013-10-01 at 14:14 +0330, Hooman Valibeigi wrote:
> I understand the prime of challenge/response protocol. Failing the
> first request looks fine as long as it occurs only once and not for
> every page you visit.
>
> I wonder if administrators would be happy with the fact that users
> have to send 2 requests to fetch an object, 40% of times on a browser
> that's been open for the whole day. Could I blame the browser for not
> learning how it should talk the proxy?
>
> Apart from the waste of bandwidth (although negligible), the other
> problem is that logs will be cluttered and full of garbage which also
> makes access/usage statistics inaccurate.

acl AuthRequest http_status 407
access_log ... !AuthRequest ...

i use the above to keep auth requests out of my logs. i dont care to
see that expected behavior is happening (given the rate at which it
happens and how clouded the logs become when 407s are logged).

i work with another proxy vendor in my job and they have functionality
to cache the users credentials for a given period of time and use
something like the users IP as a surrogate credential for the given
credential cache period. i have not dug into the authentication helpers
for squid that deeply, but do they have a similar functionality?

as for failing on the first request, the browser does not learn that it
has to provide auth, and does not begin doing so on its own. the
challenge for auth is always a function of the proxy and a browser does
not assume it will always face an auth challenge.

as an anecdotal argument, at work i recently went though a migration
from NTLM auth only to Kerberos auth as primary with NTLM as secondary
auth. because we were using an external device to handle the NTLM
challenges between the proxies and AD, along with a couple of other
performance-sapping, latency-introducing options, we significantly
impacted browsing with this configuration.

NTLM is very chatty, and authenticating every single object that an HTTP
method is issued for (GET, or whatever) means the "cost" of browsing
goes up significantly. we moved to Kerberos which reduces load on the
proxies and AD, we now have the proxies talking directly to AD instead
of the external device to speed up the auth process overall, and we
leveraged the credential caching functionality to further reduce load
and quicken the overall user experience when browsing the web.
Received on Tue Oct 01 2013 - 11:48:02 MDT

This archive was generated by hypermail 2.2.0 : Tue Oct 01 2013 - 12:00:04 MDT