Open Stack

On 21.9.2017 12:31, Giulio Fidente wrote:
> On 09/20/2017 07:36 PM, James Slagle wrote:
>> On Tue, Sep 19, 2017 at 8:37 AM, Giulio Fidente <gfidente at redhat.com> wrote:
>>> On 09/18/2017 05:37 PM, James Slagle wrote:
>>>> - The entire sequence and flow is driven via Mistral on the Undercloud
>>>> by default. This preserves the API layer and provides a clean reusable
>>>> interface for the CLI and GUI.
>>>
>>> I think it's worth saying that we want to move the deployment steps out
>>> of heat and in ansible, not in mistral so that mistral will run the
>>> workflow only once and let ansible go through the steps
>>>
>>> I think having the steps in mistral would be a nice option to be able to
>>> rerun easily a particular deployment step from the GUI, versus having
>>> them in ansible which is instead a better option for CLI users ... but
>>> it looks like having them in ansible is the only option which permits us
>>> to reuse the same code to deploy an undercloud because having the steps
>>> in mistral would require the undercloud installation itself to depend on
>>> mistral which we don't want to
>>>
>>> James, Dan, please comment on the above if I am wrong
>>
>> That's correct. We don't want to require Mistral to install the
>> Undercloud. However, I don't think that necessarily means it has to be
>> a single call to ansible-playbook. We could have multiple invocations
>> of ansible-playbook. Both Mistral and CLI code for installing the
>> undercloud could handle that easily.
>>
>> You wouldn't be able to interleave an external playbook among the
>> deploy steps however. That would have to be done under a single call
>> to ansible-playbook (at least how that is written now). We could
>> however have hooks that could serve as integration points to call
>> external playbooks after each step.
> 
> the benefits of driving the steps from mistral are that then we could
> also interleave the deployment steps and we won't need the
> ansible-playbook hook for the "external" services:
> 
> 1) collect the ansible tasks *and* the workflow_tasks (per step) from heat
> 
> 2) launch the stepN deployment workflow (ansible-playbook)
> 
> 3) execute any workflow_task defined for stepN (like ceph-ansible playbook)
> 
> 4) repeat 2 and 3 for stepN+1
> 
> I think this would also provide a nice interface for the UI ... but then
> we'd need mistral to be able to deploy the undercloud
> 

Alternatively we could do the main step loop in Ansible directly, and 
have the tasks do whatever they need to get the particular service 
deployed, from  to launching a nested ansible-playbook run if that's 
what it takes.

That way we could run the whole thing end-to-end via ansible-playbook, 
or if needed one could execute smaller bits by themselves (steps or 
nested playbook runs) -- that capability is not baked in by default, but 
i think we could make it so.

Also the interface for services would be clean and simple -- it's always 
the ansible tasks.

And Mistral-less use cases become easier to handle too (= undercloud 
installation when Mistral isn't present yet, or development envs when 
you want to tune the playbook directly without being forced to go 
through Mistral).

Logging becomes a bit more unwieldy in this scenario though, as for the 
nested ansible-playbook execution, all output would go into a task in 
the outer playbook, which would be harder to follow and the log of the 
outer playbook could be huge.

So this solution is no silver bullet, but from my current point of view 
it seems a bit less conceptually foreign than using Mistral to provide 
step loop functionality to Ansible, which should be able to handle that 
on its own.

>>>> - It would still be possible to run ansible-playbook directly for
>>>> various use cases (dev/test/POC/demos). This preserves the quick
>>>> iteration via Ansible that is often desired.
>>>>
>>>> - The remaining SoftwareDeployment resources in tripleo-heat-templates
>>>> need to be supported by config download so that the entire
>>>> configuration can be driven with Ansible, not just the deployment
>>>> steps. The success criteria for this point would be to illustrate
>>>> using an image that does not contain a running os-collect-config.
>>>>
>>>> - The ceph-ansible implementation done in Pike could be reworked to
>>>> use this model. "config download" could generate playbooks that have
>>>> hooks for calling external playbooks, or those hooks could be
>>>> represented in the templates directly. The result would be the same
>>>> either way though in that Heat would no longer be triggering a
>>>> separate Mistral workflow just for ceph-ansible.
>>>
>>> I'd say for ceph-ansible, kubernetes and in general anything else which
>>> needs to run with a standard playbook installed on the undercloud and
>>> not one generated via the heat templates... these "external" services
>>> usually require the inventory file to be in different format, to
>>> describe the hosts to use on a per-service basis, not per-role (and I
>>> mean tripleo roles here, not ansible roles obviously)
>>>
>>> About that, we discussed a more long term vision where the playbooks
>>> (static data) needd to describe how to deploy/upgrade a given service is
>>> in a separate repo (like tripleo-apb) and we "compose" from heat the
>>> list of playbooks to be executed based on the roles/enabled services; in
>>> this scenario we'd be much closer to what we had to do for ceph-ansible
>>> and I feel like that might finally allow us merge back the ceph
>>> deployment (or kubernetes deployment) process into the more general
>>> approach driven by tripleo
>>>
>>> James, Dan, comments?
>>
>> Agreed, I think this is the longer term plan in regards to using
>> APB's, where everything consumed is an external playbook/role.
>>
>> We definitely want to consider this plan in parallel with the POC work
>> that Flavio is pulling together and make sure that they are aligned so
>> that we're not constantly reworking the framework.
>>
>> I've not yet had a chance to review the material he sent out this
>> morning, but perhaps we could work together to update the sequence
>> diagram to also have a "future" state to indicate where we are going
>> and what it would look like with APB's and external paybooks.

Indeed that would be great :) IIUC, APBs are deployed by running a 
short-lived container with Ansible inside, which then connects to 
Kubernetes endpoint to create resources. So this should be a less 
complicated case than running non-containerized external playbooks.

> 
> this would be awesome, note that it isn't only ceph and kubernetes
> anymore in this scenario ... I just spotted a submission for the Skydive
> composable service and it uses the same mistral/ansible-playbook
> approach ... so it's already 3 looking forward for this!
> 
> https://review.openstack.org/#/c/502353/
> 

[1] 
https://github.com/ansibleplaybookbundle/ansible-playbook-bundle/blob/master/docs/design.md#deploy