Andrew Laski wrote:
On Wed, Sep 21, 2016, at 03:18 PM, Joshua Harlow wrote:
Andrew Laski wrote:
On Wed, Sep 21, 2016, at 12:02 PM, Joshua Harlow wrote:
Andrew Laski wrote:
However, I have asked twice now on the review what the benefit of doing
this is and haven't received a response so I'll ask here. The proposal
would add additional latency to nearly every API operation in a service
and in return what do they get? Now that it's possible to register sane
policy defaults within a project most operators do not even need to
think about policy for projects that do that. And any policy changes
that are necessary are easily handled by a config management system.
I would expect to see a pretty significant benefit in exchange for
moving policy control out of Nova, and so far it's not clear to me what
that would be.
One way to do this is to setup something like etc.d or zookeeper and
have policy files be placed into certain 'keys' in there by keystone,
then consuming projects would 'watch' those keys for being changed (and
get notified when they are changed); the project would then reload its
policy when the other service (keystone) write a new key/policy.
https://coreos.com/etcd/docs/latest/api.html#waiting-for-a-change
or
https://zookeeper.apache.org/doc/r3.4.5/zookeeperProgrammers.html#ch_zkWatches
or (pretty sure consul has something similar),
This is pretty standard stuff folks :-/ and it's how afaik things like
https://github.com/skynetservices/skydns work (and more), and it would
avoid that 'additional latency' (unless the other service is adjusting
the policy key every millisecond, which seems sorta unreasonable).
Sure. Or have Keystone be a frontend for ansible/puppet/chef/.... What's
not clear to me in any of this is what's the benefit to having Keystone
as a fronted to policy configuration/changes, or be involved in any real
way with authorization decisions? What issue is being solved by getting
Keystone involved?
I don't understand the puppet/chef connection, can u clarify.
If I'm interpreting it right, I would assume it's the same reason that
something like 'skydns' exists over etcd; to provide a useful API that
focuses on the dns particulars that etcd will of course not have any
idea about. So I guess the keystone API could(?)/would(?) then focus on
policy particulars as its value-add.
Maybe now I understand what u mean by puppet/chef, in that you are
asking why isn't skydns (for example) just letting/invoking
puppet/chef/ansible to distribute/send-out dns (dnsmasq) files? Is that
your equivalent question?
I'm focused on Nova/Keystone/OpenStack here, I'm sure skydns has good
reasons for their technical choices and I'm in no place to question
them.
I'm trying to understand the value-add that Keystone could provide here.
Policy configuration is fairly static so I'm not understanding the
desire to put an API on top of it. But perhaps I'm missing the use case
here which is why I've been asking.
My ansible/puppet/chef comparison was just that those are ways to
distribute static files and would work just as well as something built
on top of etcd/zookeeper. I'm not really concerned about how it's
implemented though. I'm just trying to understand if the desire is to
have Keystone handle this so that deployers don't need to work with
their configuration management system to configure policy files, or is
there something more here?
Gotcha, thanks for explaining.
I'll let others comment, but my semi-useful/semi-baked thoughts around
this are as a user I would want to:
#1 Can I query keystone (or perhaps I should ask nova) to show me what
(all the) APIs in nova I'm allowed to call (without actually having to
perform those same calls to figure it out); ie, tell me how my known
role/user/tenant in <something> maps to the policy stored (somewhere in
some project) so I can make smart decisions around which APIs I can be
calling.
#2 Can I go to one place (the same place that has my roles and tenants
and such?) and ensure that by changing roles or such that dependent
systems that may have meanings for those roles are not adversely
affected (or say makes a policy being used become invalid, ie similar to
a error saying 'the change you have requested violates the rules defined
in 'nova policy' and therefore is invalid and can't be applied').
The #1 kind of use-case of course would be really easy if keystone has
the knowledge of each projects 'policy.json' (or equivalent data
structure); and since keystone already has the role/user/tenant
information it would be straightforward to solve #2 there as well
(because keystone could reject a role change or tenancy change or user
change or ... if it negatively/violates affects some projects policy).
Of course if u distribute the policy information then each project would
have to implement #1 and there would need to be some 2 way mechanism to
ensure #2 happens correctly (because if keystone just blindly does
role/user/tenant changes it may violate a projects policy definition).
Just my current thoughts, I'm sure there are other thoughts around
distributing vs centralizing and so-on (just hoping that we can think
past the view that centralizing doesn't really have to imply that
keystone has to be called for every single REST API called as a
precursor, if we use systems like etcd/zookeeper/... smartly).
-Josh
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
