Open Stack

On Mon, Apr 18, 2016 at 5:43 PM, Matt Fischer <matt at mattfischer.com> wrote:

> On Mon, Apr 18, 2016 at 12:52 PM, Morgan Fainberg <
> morgan.fainberg at gmail.com> wrote:
>
>>
>>
>> On Mon, Apr 18, 2016 at 7:29 AM, Brant Knudson <blk at acm.org> wrote:
>>
>>>
>>>
>>> On Fri, Apr 15, 2016 at 9:04 PM, Adam Young <ayoung at redhat.com> wrote:
>>>
>>>> We all want Fernet to be a reality.  We ain't there yet (Except for
>>>> mfish who has no patience) but we are getting closer.  The goal is to get
>>>> Fernet as the default token provider as soon as possible. The review to do
>>>> this has uncovered a few details that need to be fixed before we can do
>>>> this.
>>>>
>>>> Trusts for V2 tokens were not working correctly.  Relatively easy fix.
>>>> https://review.openstack.org/#/c/278693/ Patch is still failing on
>>>> Python 3.  The tests are kindof racy due to the revocation event 1 second
>>>> granularity.  Some of the tests here have A sleep (1) in them still, but
>>>> all should be using the time control aspect of the unit test fixtures.
>>>>
>>>> Some of the tests also use the same user to validate a token as that
>>>> have, for example, a role unassigned.  These expose a problem that the
>>>> revocation events are catching too many tokens, some of which should not be
>>>> treated as revoked.
>>>>
>>>> Also, some of the logic for revocation checking has to change. Before,
>>>> if a user had two roles, and had one removed, the token would be revoked.
>>>> Now, however, the token will validate successful, but the response will
>>>> only have the single assigned role in it.
>>>>
>>>>
>>>> Python 3 tests are failing because the Fernet formatter is insisting
>>>> that all project-ids be valid UUIDs, but some of the old tests have "FOO"
>>>> and "BAR" as ids.  These either need to be converted to UUIDS, or the
>>>> formatter needs to be more forgiving.
>>>>
>>>> Caching of token validations was messing with revocation checking.
>>>> Tokens that were valid once were being reported as always valid. Thus, the
>>>> current review  removes all caching on token validations, a change we
>>>> cannot maintain.  Once all the test are successfully passing, we will
>>>> re-introduce the cache, and be far more aggressive about cache invalidation.
>>>>
>>>> Tempest tests are currently failing due to Devstack not properly
>>>> identifying Fernet as the default token provider, and creating the Fernet
>>>> key repository.  I'm tempted to just force devstack to always create the
>>>> directory, as a user would need it if they ever switched the token provider
>>>> post launch anyway.
>>>>
>>>>
>>> There's a review to change devstack to default to fernet:
>>> https://review.openstack.org/#/c/195780/ . This was mostly to show that
>>> tempest still passes with fernet configured. It uncovered a couple of test
>>> issues (similar in nature to the revocation checking issues mentioned in
>>> the original note) that have since been fixed.
>>>
>>> We'd prefer to not have devstack overriding config options and instead
>>> use keystone's defaults. The problem is if fernet is the default in
>>> keystone then it won't work out of the box since the key database won't
>>> exist. One option that I think we should investigate is to have keystone
>>> create the key database on startup if it doesn't exist.
>>>
>>>
>> I am unsure if this is the right path, unless we consider possibly moving
>> the key-DB for fernet into the SQL backend (possible?) notably so we can
>> control a cluster of keystones.
>>
>
> -1 this idea. Now we've got to figure out how to lock the DB and where
> will those keys/passwords go? The current model is apparently painful for
> devstack,  but it works fine for anyone who has an automated deployment
> system, or something else simple like rsync/scp can be used. Fernet keys
> are simpler to update/replicate than PKI certs and we didn't store those in
> the DB to my knowledge. Let's please not optimize this solution for
> devstack.
>
>
The DB solution is actually presented as explicitly not AIO install, it
would be to solve the "cluster case" as the default value for keystone's
config, since keystone can reliably share the data. DB tables can actually
be locked (this is separate from select for update) and would be less
painful.

The Devstack solution today (and looking to the future) would absolutely be
file-based. No change there. This conversation is revolving around how to
reduce the operational overhead of making keystone itself (not in
devstack's context) default to fernet and not produce wildly (or
potentially wildly) broken states.

This is a convo that will need to occur at the summit and we can hash out
the details of what makes sense more easily. I expect the DB solution would
be a bad one in any case, but before we make keystone's config default to
Fernet, we need to make the non-devstack case(s) (and existing deployments)
not suck.

--Morgan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20160418/82e91c4c/attachment.html>