Open Stack

Excerpts from Gage Hugo's message of 2017-12-19 11:14:18 -0600:
> On Tue, Dec 12, 2017 at 5:34 PM, Doug Hellmann <doug at doughellmann.com>
> wrote:
> 
> > Excerpts from Dave McCowan (dmccowan)'s message of 2017-12-12 21:36:51
> > +0000:
> > >
> > > On 12/12/17, 3:15 PM, "Doug Hellmann" <doug at doughellmann.com> wrote:
> > >
> > > >Excerpts from Dave McCowan (dmccowan)'s message of 2017-12-12 19:56:49
> > > >+0000:
> > > >>
> > > >> On 12/12/17, 10:38 AM, "Doug Hellmann" <doug at doughellmann.com> wrote:
> > > >>
> > > >> >
> > > >> >> On Dec 12, 2017, at 9:42 AM, Paul Bourke <paul.bourke at oracle.com>
> > > >>wrote:
> > > >> >>
> > > >> >> From my understanding it would be a cleanup operation - which to be
> > > >> >>honest, would be very much welcomed. I recently did a little work
> > with
> > > >> >>Castellan to integrate it with Murano and found the auth code to be
> > > >>very
> > > >> >>messy, and flat out broken in some cases. If it's possible to let
> > the
> > > >> >>barbican client take care of this that sounds good to me.
> > > >> >>
> > > >> >> > Which mode is used the most in the services that consume
> > castellan
> > > >> >> > today?
> > > >> >>
> > > >> >> Afaik Barbican is the only backend that currently exists in
> > Castellan
> > > >> >>[0]. Looking again it seems some support has been added for vault
> > > >>which
> > > >> >>is great, but I reckon Barbican would still be the primary use.
> > > >> >>
> > > >> >> I haven't been hugely active in Castellan but if the team would
> > like
> > > >> >>some more input on this or reviews please do ping me, I'd be glad to
> > > >> >>help.
> > > >> >
> > > >> >What I mean is, in the services consuming Castellan, how do they
> > expect
> > > >> >it to authenticate to Barbican? As the current user or as a
> > hard-coded
> > > >> >fixed user controlled by the deployer? I would think most services
> > > >>would
> > > >> >need to connect as the ³current² user talking to them so they can
> > > >>access
> > > >> >that user¹s secrets from Barbican. Removing the keystoneauth stuff
> > from
> > > >> >the driver would therefore break all of those applications.
> > > >> >
> > > >> >Doug
> > > >>
> > > >> We're a mix right now.  Nova and Cinder pass through the a user's
> > token
> > > >>to
> > > >> retrieve the user's key for encrypted volumes.  Octavia uses its
> > service
> > > >> account to retrieve certificates for load balancing TLS connections.
> > > >> Users must grant Octavia read permissions in advance.
> > > >
> > > >OK, so it sounds like we do need to continue to support both
> > > >approaches to authentication.
> > > >
> > > >> Keystone is currently the only authentication option for Barbican.  I
> > > >> believe the proposal to decouple keystoneauth is advance work for
> > adding
> > > >> new auth methods and backends as future work.  Vault and Custodia are
> > > >>two
> > > >> such backends in progress.  They don't support keystoneauth and likely
> > > >> won't, so we'll need alternatives.
> > > >
> > > >Each driver manages its own authentication, right? Why do we need to
> > > >remove the keystoneauth stuff in the barbican driver in order to enable
> > > >other drivers?
> > >
> > > I would use the word "decouple", with the intent to give the option of
> > > using Castellan without having a dependency on keystoneauth.  But, I
> > don't
> > > want to speak for original posters who used the word "remove" in case
> > they
> > > have other ideas.
> > >
> > > Until recently Barbican was the only secret store and Keystone was the
> > > only authentication service, so we didn't have to sort through the
> > > modularity.
> >
> > I'm sorry that I missed the conversation about this in Denver.  It
> > seems like everyone else understands what's being proposed in a way
> > very different than I do, so I apologize for continuing to just ask
> > the same questions.  I'll try rephrasing, but it would be *very*
> > helpful if someone would summarize that discussion and lay out the
> > plan in more detail than "we want to remove the use of keystoneauth."
> > If we can't do it by email, then maybe via an Oslo spec.
> >
> >
> > The barbican driver has 2 modes for using keystoneauth. One is to
> > use the use the execution context to authenticate using the same
> > token that the current user passed in to the service calling into
> > castellan. The other is to use credentials from the configuration
> > file.
> >
> > Those options seem to be pretty well abstracted in the API, so that
> > the application using castellan can just pass the right sort of
> > context and get the right behavior from the driver, without having
> > to know what the driver is. We currently only have a barbican driver,
> > and that driver uses keystoneauth directly because that is the only
> > way to control which authentication mode is used. Other drivers
> > would presumably use some means other than keystoneauth to authenticate
> > to the backends they talk to, with the difference in behavior (acting
> > like the current user or acting like a service user) triggered by
> > the context passed in.
> >
> > If we don't use keystoneauth inside the castellan driver before
> > creating the barbican client, how will we support both modes in the
> > castellan API without exposing to the application which secret store
> > driver is being used? We can't, for example, require that an
> > application configured to use the barbican driver pass more (or
> > different) information to castellan than it would pass if castellan
> > was configured to use custodia, because that would break the
> > abstraction.
> >
> 
> I wonder if we could make keystoneauth a soft requirement instead for those
> using
> the Barbican driver as a way to de-couple it?  Then if one were to use a
> different
> backend (Vault/Custodia/etc.) it wouldn't be needed.
> 
> Not sure how having different backends will be though
> (Barbican/Vault/Custodia) in terms
> of breaking abstraction.

I hope it doesn't break anything. The point of Castellan was to provide
that abstraction, right? :-)

We could work on making keystoneauth a driver-specific requirement for
castellan during rocky. Effectively it's not going to make much
difference, because currently everything that uses casetellan also
relies on talking to keystone for other reasons so keystoneauth is still
going to need to be installed. But we can adjust the requirements using
the "extras" feature of setuptools to clarify which drivers use each
dependency.

> 
> >
> > Are there more extensive changes planned for the public API of
> > castellan, to use different mechanisms to get a driver handle for
> > the different modes?  Given our backwards-compatibility constraints,
> > we can't change the API of the library in a breaking way without
> > also updating the consuming apps, so we would have to *add* an API
> > and deprecate the old one. I haven't seen anyone talk about a new
> > API, though.
> 
> > Are we planning to drop support for one access mode, and change the
> > way castellan works more fundamentally? This possibility raises the
> > same questions as changing the API does.  Based on the compatibility
> > constraints for an Oslo library, we need to continue to support
> > both of modes until we are sure they are not being used by any of
> > our applications.
> >
> 
> I'm not sure about these, maybe someone on the Castellan team could chime
> in here.
> 
> >
> > Doug
> >
> > __________________________________________________________________________
> > OpenStack Development Mailing List (not for usage questions)
> > Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> >