Open Stack

Top posting just to try and summarize my thought that for the goal in 
Stein, I think we should focus on getting the base framework in place 
for each service project, along with any non-config (including policy) 
specific upgrade checks that make sense for each project.

As Ben mentioned, there are existing tools for validating config (I know 
BlueBox used to use the fatal_deprecations config in their CI/CD 
pipeline to know when they needed to change their deploy scripts because 
deploying new code from pre-prod would fail). Once we get the basics 
covered we can work, as a community, to figure out how best to integrate 
config validation into upgrade checks, because I don't really think we 
want to have upgrade checks that dump warnings for all deprecated 
options in addition to what is already provided by oslo.config/log. I 
have a feeling that would get so noisy that no one would ever pay 
attention to it. I'm mostly interested in the scenario that config is 
removed from code but still being set in the config file which could 
fail an upgrade on service restart (if an alias was removed for 
example), but I also tend to think those types of issues are case-by-case.

On 10/15/2018 3:29 PM, Ben Nemec wrote:
> 
> 
> On 10/15/18 3:27 AM, Jean-Philippe Evrard wrote:
>> On Fri, 2018-10-12 at 17:05 -0500, Matt Riedemann wrote:
>>> The big update this week is version 0.1.0 of oslo.upgradecheck was
>>> released. The documentation along with usage examples can be found
>>> here
>>> [1]. A big thanks to Ben Nemec for getting that done since a few
>>> projects were waiting for it.
>>>
>>> In other updates, some changes were proposed in other projects [2].
>>>
>>> And finally, Lance Bragstad and I had a discussion this week [3]
>>> about
>>> the validity of upgrade checks looking for deleted configuration
>>> options. The main scenario I'm thinking about here is FFU where
>>> someone
>>> is going from Mitaka to Pike. Let's say a config option was
>>> deprecated
>>> in Newton and then removed in Ocata. As the operator is rolling
>>> through
>>> from Mitaka to Pike, they might have missed the deprecation signal
>>> in
>>> Newton and removal in Ocata. Does that mean we should have upgrade
>>> checks that look at the configuration for deleted options, or
>>> options
>>> where the deprecated alias is removed? My thought is that if things
>>> will
>>> not work once they get to the target release and restart the service
>>> code, which would definitely impact the upgrade, then checking for
>>> those
>>> scenarios is probably OK. If on the other hand the removed options
>>> were
>>> just tied to functionality that was removed and are otherwise not
>>> causing any harm then I don't think we need a check for that. It was
>>> noted that oslo.config has a new validation tool [4] so that would
>>> take
>>> care of some of this same work if run during upgrades. So I think
>>> whether or not an upgrade check should be looking for config option
>>> removal ultimately depends on the severity of what happens if the
>>> manual
>>> intervention to handle that removed option is not performed. That's
>>> pretty broad, but these upgrade checks aren't really set in stone
>>> for
>>> what is applied to them. I'd like to get input from others on this,
>>> especially operators and if they would find these types of checks
>>> useful.
>>>
>>> [1] https://docs.openstack.org/oslo.upgradecheck/latest/
>>> [2] https://storyboard.openstack.org/#!/story/2003657
>>> [3]
>>> http://eavesdrop.openstack.org/irclogs/%23openstack-dev/%23openstack-dev.2018-10-10.log.html#t2018-10-10T15:17:17 
>>>
>>> [4]
>>> http://lists.openstack.org/pipermail/openstack-dev/2018-October/135688.html 
>>>
>>>
>>
>> Hey,
>>
>> Nice topic, thanks Matt!
>>
>> TL:DR; I would rather fail explicitly for all removals, warning on all
>> deprecations. My concern is, by being more surgical, we'd have to
>> decide what's "not causing any harm" (and I think deployers/users are
>> best to determine what's not causing them any harm).
>> Also, it's probably more work to classify based on "severity".
>> The quick win here (for upgrade-checks) is not about being smart, but
>> being an exhaustive, standardized across projects, and _always used_
>> source of truth for upgrades, which is complemented by release notes.
>>
>> Long answer:
>>
>> At some point in the past, I was working full time on upgrades using
>> OpenStack-Ansible.
>>
>> Our process was the following:
>> 1) Read all the project's releases notes to find upgrade documentation
>> 2) With said release notes, Adapt our deploy tools to handle the
>> upgrade, or/and write ourselves extra documentation+release notes for
>> our deployers.
>> 3) Try the upgrade manually, fail because some release note was missing
>> x or y. Find root cause and retry from step 2 until success.
>>
>> Here is where I see upgrade checkers improving things:
>> 1) No need for deployment projects to parse all release notes for
>> configuration changes, as tooling to upgrade check would be directly
>> outputting things that need to change for scenario x or y that is
>> included in the deployment project. No need to iterate either.
>>
>> 2) Test real deployer use cases. The deployers using openstack-ansible
>> have ultimate flexibility without our code changes. Which means they
>> may have different code paths than our gating. Including these checks
>> in all upgrades, always requiring them to pass, and making them
>> explicit about the changes is tremendously helpful for deployers:
>> - If config deprecations are handled as warnings as part of the same
>> process, we will output said warnings to generate a list of action
>> items for the deployers. We would use only one tool as source of truth
>> for giving the action items (and still continue the upgrade);
>> - If config removals are handled as errors, the upgrade will fail,
>> which is IMO normal, as the deployer would not have respected its
>> action items.
> 
> Note that deprecated config opts should already be generating warnings 
> in the logs. It is also possible now to use fatal-deprecations with 
> config opts: 
> https://github.com/openstack/oslo.config/commit/5f8b0e0185dafeb68cf04590948b9c9f7d727051 
> 
> 
> I'm not sure that's exactly what you're talking about, but those might 
> be useful to get us at least part of the way there.
> 
>>
>> In OSA, we could probably implement a deployer override (variable). It
>> would allow the deployers an explicit bypass of an upgrade failure. "I
>> know I am doing this!". It would be useful for doing multiple serial
>> upgrades.
>>
>> In that case, deployers could then share together their "recipes" for
>> handling upgrade failure bypasses for certain multi-upgrade (jumps)
>> scenarios. After a while, we could think of feeding those back to
>> upgrade checkers.
>>
>> 3) I like the approach of having oslo-config-validator. However, I must
>> admit it's not part of our process to always validate a config file
>> before trying to start a service in OSA. I am not sure where other
>> deployment projects are in terms of that usage. I am not familiar with
>> upgrade checker code, but I would love to see it re-using oslo-config-
>> validator, as it would be the unique source of truth for upgrades
>> before the upgrade happens (vs having to do multiple steps).
>> If I am completely out of my league here, tell me.
> 
> This is a bit tricky as the validator requires information that is not 
> necessarily available in a production environment. Specifically, it 
> either needs the oslo-config-generator configuration file that lists all 
> of the namespaces a project uses, or it needs a generated 
> machine-readable sample config that contains all of the opt data. The 
> latter is not generally available today, and I'm not sure whether the 
> former is either. A quick pip install of an OpenStack service suggests 
> that it is not.
> 
> Ideally, the machine-readable sample config would be available from 
> packages anyway as it has other uses too, but it's a pretty big ask to 
> get all of the packagers shipping that this cycle. I'm not sure how it 
> would work with pip installs either, although it seems like we should be 
> able to figure out something there.
> 
> Anyway, not saying we shouldn't do it, but I want to make it clear that 
> this isn't as simple as just adding one more check to the upgrade 
> checkers. There are some other dependencies to doing this in a 
> non-service-specific way.
> 
>>
>> Just my 2 cents.
>> Jean-Philippe Evrard (evrardjp)
>>
>>
>>
>> _______________________________________________
>> OpenStack-operators mailing list
>> OpenStack-operators at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>
> 
> __________________________________________________________________________
> 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


-- 

Thanks,

Matt