On 03/16/2015 03:24 PM, Doug Hellmann wrote:
Excerpts from Adam Young's message of 2015-03-16 14:17:16 -0400:
On 03/16/2015 01:45 PM, Doug Hellmann wrote:
All of these are reasons we have so far resisted building a service to
deploy updates to oslo.config's input files, and rely on provisioning
tools to update them.
Have we consider using normal provisioning tools for pushing out
changes to policy files, and having the policy library look at the
timestamp of the file(s) to decide if it needs to re-read them
before evaluating a rule? Maybe we wouldn't always scan the file
system, but wait for some sort of signal that the scan needs to be
done.
I like this last idea. Thew trigger needs to be app specific, I think.
I was thinking of a callback to be triggered by 'kill -HUP $pid'. We can
make a little framework for registering callbacks on signals (if there
isn't something like that already) to allow multiple refresh actions on
the signal.
Doug
I think policy files are not config files. We've treated them as such
in the past as they are not dynamic, but I don't think I want to *have*
to do this:
1. Change policy in keystone (somehow)
2. Tell Puppet that there is a new file
3. Have puppet pick up the3 new file and sync it to the servers.
Right, I wouldn't do that. I would modify the file in my puppet
repository and then push that out all at once. Keystone would receive
the policy files the same way as the other services.
Although I would say that we should make it easy to support this workflow.
For one thing, it assumes that all of the comsuers are talking to the
same config management system, which is only true for a subset of the
services.
I'm not sure what you mean here. Do you mean that in a given deployment
you would expect some services to be configured by puppet and others to
be configured a different way?
I mean that there could be a puppet server for the core infrastructure,
and another one (or Ansible or Chef) for Hadoop on top of that. There
is no one puppet master that we can assume to be controlling all of the
servers. They might be run by different organizations.
I see a case for doing this same kind of management for Many of the
files Keystone produces. Service catalog is the most obvious candidate.
Yes, that's another good example, although in that case we do already
have an API that lets a cloud consumer access the service catalog data
so it might be viewed as different from the policy rules or oslo.config
files (the latter at least typically have private data we wouldn't want
to share through an API).
We don't have a single, monolithic Service catalog (anymore) and, with
endpoint filtering, we expect multiple service catalogs to be the norm.
I want to pursue the idea of git style file identification here, (hash
of the file as identifier) as that works to split the service catalog
from the token, and still have multiple service catalogs, but ensure
that they are correctly linked in remote systems. It doesn't have to
be hash, but it makes the process much more verifiable. This is also
true for policy files; there can be more than one active at any given
point in time, fetchable by remote identifier. Even as we push towards
common rules for defininng the RBAC section, we have to be aware that
different endpoints might need different policy files.
If we could have a workflow for managing : PKI certs, Federatiomn
mappings and (Group only?) Role Assignments we could decentralize token
validation.
When doing the PKI tokens, we discussed this, and ended up with a t
"fetch first" policy toward the certs.
Puppet does not know how to get a token, so it can't call the keystone
token-protected APIs to fetch new data. What forms of authentication do
the config managment systems support? Is this an argument for tokenless
operations against Keystone?
In my scenario puppet (or chef or whatever) is the source of truth for
the configuration file, not one of our services. So there's no need for
the configuration management tool to talk to any of our services beyond
sending the HUP signal telling us to re-read the file(s).
So services would generate files to be published to Puppet. As I said,
that would work for a subset of use cases, and probably makes sense for
core infrastructure, but was cannot assume all consumers are talking to
puppet, or even if they are, talking to the same puppet master.
Doug
__________________________________________________________________________
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
__________________________________________________________________________
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
