On 05/21/2014 08:23 PM, John Dickinson wrote:
On May 21, 2014, at 4:26 PM, Adam Young <ayo...@redhat.com> wrote:
On 05/21/2014 03:36 PM, Kurt Griffiths wrote:
Good to know, thanks for clarifying. One thing I'm still fuzzy on, however, is
why we want to deprecate use of UUID tokens in the first place? I'm just trying
to understand the history here...
Because they are wasteful, and because they are the chattiest part of
OpenStack. I can go into it in nauseating detail if you really want, including
the plans for future enhancements and the weaknesses of bearer tokens.
A token is nothing more than a snap shot of the data you get from Keystone
distributed. It is stored in Memcached and in the Horizon session uses the
hash of it for a key.
You can do the same thing. Once you know the token has been transferred once
to a service, assuming that service has caching on, you can pass the hash of
the key instead of the whole thing.
So this would mean that a Swift client would auth against Keystone to get the PKI token,
send that to Swift, and then get back from Swift a "short" token that can be
used for subsequent requests? It's an interesting idea to consider, but it is a new sort
of protocol for clients to implement.
It would probably be more correct for Swift to calculate that, yes, but
the client could also just calculate the hash and send it on subsequent
requests. As you pointed out, it is a matter of performance.
Actually, you can do that up front, as auth_token middleware will just default
to an online lookup. However, we are planning on moving to ephemeral tokens
(not saved in the database) and an online lookup won't be possible with those.
The people that manage Keystone will be happy with that, and forcing an online
lookup will make them sad.
An "online lookup" is one that calls the Keystone service to validate a token?
Which implies that by disabling online lookup there is enough info in the token to
validate it without any call to Keystone?
Yes. the whole popen call to openssl to verify the messages.
I understand how it's advantageous to offload token validation away from Keystone itself
(helps with scaling), but the current "solution" here seems to be pushing a lot
of pain to consumers and deployers of data APIs (eg Marconi and Swift and others).
We try to encapsulate it all within auth_token middleware, but the
helper functions are in python-keystoneclient if you need more specific
handling.
Hash is MD5 up through what is released in Icehouse. The next version of
auth_token middleware will support a configurable algorithm. The default
should be updated to sha256 in the near future.
If a service (like Horizon) is hashing the token and using that as a session
key, then why does it matter what the auth_token middleware supports? Isn't the
hashing handled in the service itself? I'm thinking in the context of how we
would implement this idea in Swift (exploring possibilities, not committing to
a patch).
That is after it has received the token. So, Horizon could send the
hash to Nova, and Nova would then be required to make the call to
Keystone, just like UUID tokens. That would break on the ephemeral
approach.
I'm exploring the Horizon side of the equasion for some other reasons,
primarily in the context of Kerberos support, but also for better
revocation rules. If the onus is on the client (in this case Horizon)
to remember if it has send a particular token in full form it might be a
little hard to keep track.
What communication is most impacted by the large token size? Is it
fetching out images for a web page or something like that?
From: Morgan Fainberg <morgan.fainb...@gmail.com>
Reply-To: OpenStack Dev <openstack-dev@lists.openstack.org>
Date: Wednesday, May 21, 2014 at 1:23 PM
To: OpenStack Dev <openstack-dev@lists.openstack.org>
Subject: Re: [openstack-dev] Concerns about the ballooning size of keystone
tokens
This is part of what I was referencing in regards to lightening the data stored in the
token. Ideally, we would like to see an "ID only" token that only contains the
basic information to act. Some initial tests show these tokens should be able to clock in
under 1k in size. However all the details are not fully defined yet. Coupled with this
data reduction there will be explicit definitions of the data that is meant to go into
the tokens. Some of the data we have now is a result of convenience of accessing the data.
I hope to have this token change available during Juno development cycle.
There is a lot of work to be done to ensure this type of change goes smoothly.
But this is absolutely on the list of things we would like to address.
Cheers,
Morgan
Sent via mobile
On Wednesday, May 21, 2014, Kurt Griffiths <kurt.griffi...@rackspace.com> wrote:
adding another ~10kB to each request, just to save a once-a-day call to
Keystone (ie uuid tokens) seems to be a really high price to pay for not
much benefit.
I have the same concern with respect to Marconi. I feel like KPI tokens
are fine for control plane APIs, but don't work so well for high-volume
data APIs where every KB counts.
Just my $0.02...
--Kurt
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev