IDP node stickiness without an SLB?

[Wessel, Keith]

FYI, we chickened out. Though Dave's solution seems workable as does the University of Washington mod_proxy solution, we didn't want to get overly complicated simply to avoid an extra VM.

We decided to go with active/passive. If this wasn't a VM setup, these other solutions would have had much more value for us. But in this environment, some extra memory or CPU is nothing.

Thanks, all, for the ideas.

Keith


From: users-bounces at shibboleth.net [mailto:users-bounces at shibboleth.net] On Behalf Of David Langenberg
Sent: Friday, January 24, 2014 4:13 PM
To: Shib Users
Subject: Re: IDP node stickiness without an SLB?

I'd assume the browser would've done the DNS query once rather than once-per-connection, but if we have to get difficult then so be it.  You could rectify it by having the login.jsp detect the IP address of the node it's running on and send down all further references for elements in the page to be grabbed by IP address rather than by DNS name.  That'd force everything to that single node.

Dave

On Fri, Jan 24, 2014 at 2:41 PM, Wessel, Keith <kwessel at illinois.edu<mailto:kwessel at illinois.edu>> wrote:
This is an interesting suggestion, Dave. And at first, seems like that'll do the trick. Enter concurrent connections. Looks like the HTTP 1.1 spec recommends browsers support two persistent connections to a given host to speed up page loading by doing it in parallel. With the launch of IE 8, our friends at Microsoft bumped their browser up to a max of six concurrent persistent connections.

Problem arises if one of those connections grabs the 1st IP from the load balancer and the next one grabs the other. Usually won't happen with DNS caching, but depending on when consecutive connections are opened, I can theoretically see a problem here.

Know any magic to work around this one? We really like this idea if we can make it work!

Keith


From: users-bounces at shibboleth.net<mailto:users-bounces at shibboleth.net> [mailto:users-bounces at shibboleth.net<mailto:users-bounces at shibboleth.net>] On Behalf Of David Langenberg
Sent: Wednesday, January 22, 2014 4:56 PM

To: Shib Users
Subject: Re: IDP node stickiness without an SLB?

You could extend that 60 second timeout by adding an AJAX request back to the IdP every, say, 10 seconds for something silly like making a HEAD request on some CSS or text file.

Dave

On Wed, Jan 22, 2014 at 3:51 PM, Wessel, Keith <kwessel at illinois.edu<mailto:kwessel at illinois.edu>> wrote:
David,

I haven't done more than research this idea, and it's an interesting one. Problem is, from what I've read, browsers are going to have their own persistent connection timeout. IE seems to default to 60 seconds, for instance. That's long enough for most users to log in, but I don't think we want a 60-second lag on the login screen to cause a user to get an error.

This is such a simple and clean solution other than this and, of course, the fact that you'll have more httpd processes lingering around.

Still considering the IDP proxy model. But all of this sure is making active/passive tempting.

Keith


From: users-bounces at shibboleth.net<mailto:users-bounces at shibboleth.net> [mailto:users-bounces at shibboleth.net<mailto:users-bounces at shibboleth.net>] On Behalf Of David Langenberg
Sent: Tuesday, January 21, 2014 4:36 PM

To: Shib Users
Subject: Re: IDP node stickiness without an SLB?

Just a complete shot in the dark, but what if you set each node to use an HTTP KeepAlive with a long timeout?  The idea being getting the browser to maintain an open connection to the node while the user is entering their credentials.  Submission of the form would then (hopefully) happen over the existing connection back to the same node & thus allow everything to succeed.

Dave

On Tue, Jan 21, 2014 at 3:21 PM, Wessel, Keith <kwessel at illinois.edu<mailto:kwessel at illinois.edu>> wrote:
All,

I'm trying to find a way for our IDP nodes to remain sticky for a user throughout the login process in a GSLB environment where client stickiness isn't an option. The GSLB talks to DNS servers, not clients, and we don't want our entire campus DNS resolver to be sticky to a single IDP node if we can help it.

We deployed Paul's excellent database-backed storage service but had to back out the change. As I understand it, the login process has to stick to a single IDP node when using this solution and, sure enough, we saw problems when a user's browser started on one IDP node then went to the other, either while getting to the login page or when the login page was submitted. Since the session wasn't yet known to the other IDP node, the login couldn't be completed.

Can anyone make any suggestions of how we might, once a user hits node 1 or node 2, keep them going back to that node until they're logged in and returning to the SP? I see no easy way to, say, put an absolute URL in place of /idp for all of the relative URLs presented by the IDP.

If we can't accomplish this, we're planning to still use the database-backed storage service combined with active-passive from our GSLB. But we really do like being able to balance traffic between the two sites when they're both available.

Thanks,
Keith


--
To unsubscribe from this list send an email to users-unsubscribe at shibboleth.net<mailto:users-unsubscribe at shibboleth.net>



--
David Langenberg
Identity & Access Management
The University of Chicago

--
To unsubscribe from this list send an email to users-unsubscribe at shibboleth.net<mailto:users-unsubscribe at shibboleth.net>



--
David Langenberg
Identity & Access Management
The University of Chicago

--
To unsubscribe from this list send an email to users-unsubscribe at shibboleth.net<mailto:users-unsubscribe at shibboleth.net>



--
David Langenberg
Identity & Access Management
The University of Chicago
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://shibboleth.net/pipermail/users/attachments/20140217/6acb6ba9/attachment-0001.html