Open Stack

On 29.10.2015 15:24, Bogdan Dobrelya wrote:
> Hello.
> There are few types of a deployment regressions possible. When changing
> a module version to be used from upstream (or internal module repo), for
> example from Liberty to Mitaka. Or when changing the composition layer
> (modular tasks in Fuel). Specifically, adding/removing/changing classes
> and a class parameters.
> 
> An example regression for swift deployment data [0]. Something was
> changed unnoticed by existing noop tests and as a result
> the swift data became being stored in root partition.
> 
> Suggested per-commit based regressions detection [1] for deployment data
> assumes to automatically detect if a class in a noop catalog run has
> gained or lost a parameter or if it has been updated to another value by
> a patch under test. Later, this check could even replace existing noop
> tests, everything will be checked automatically, unless every deployment
> scenario is covered by a corresponding template, which are represented
> as YAML files [2] in Fuel.
> Note: The tool [3] can help to get all deployment cases (-Y) and all
> deployment tasks (-S) as well.
> 
> I propose to review the patch [1], understand how it works (see tl;dr
> section below) and to start using it ASAP. The earlier we commit the
> "initial" data layer state, less regressions would pop up.
> 
> (tl;dr)
> The check should be done for every modular component (aka deployment
> task). Data generated in the noop catalog run for all classes and
> defines of a given deployment task should be verified against its
> "acknowledged" (committed) state.
> And fail the test gate, if changes has been found, like new parameter
> with a defined value, removed a parameter, changed a parameter's value.
> 
> In order to remove a regression, a patch author will have to add (and
> reviewers should acknowledge) detected changes in the committed state of
> the deployment data. This may be done manually, with a tool like [3] or
> by a pre-commit hook, or even at the CI side!
> The regression check should show the diff between committed state and a
> new state proposed in a patch. Changed state should be *reviewed* and
> accepted with a patch, to became a committed one. So the deployment data
> will evolve with *only* approved changes. And those changes would be
> very easy to be discovered for each patch under review process!
> No more regressions, everyone happy.
> 
> Examples:
> 
> - A. A patch author removed the mpm_module parameter from the
> composition layer (apache modular task). The test should fail with a
> 
> Diff:
>       @@ -90,7 +90,7 @@
>          manage_user            => 'true',
>          max_keepalive_requests => '100',
>          mod_dir                => '/etc/httpd/conf.d',
>       -  mpm_module             => 'false',
>       +  mpm_module             => 'prefork',
>          name                   => 'Apache',
>          package_ensure         => 'installed',
>          ports_file             => '/etc/httpd/conf/ports.conf',
> 
> It illustrates that the mpm_module's committed value was "false".
> But the new one came as the 'prefork', likely from the apache class
> defaults.
> The solution:
> Follow the failed build link and see for detected changes (a diff).
> Acknowledge the changes and include rebuilt templates in the patch as a
> new revision. The tool [3] (use -h for help) example command:
> ./utils/jenkins/fuel_noop_tests.rb -q -b -s api-proxy/api-proxy_spec.rb
> 
> Or edit the committed templates manually and include data changes in the
> patch as well.
> 
> -B. An upstream module author added the new parameter mpm_mode with a
> default '123'. The test should fail with a
> 
> Diff:
>        @@ -90,6 +90,7 @@
>           manage_user            => 'true',
>           max_keepalive_requests => '100',
>           mod_dir                => '/etc/httpd/conf.d',
>        +  mpm_mode               => '123',
>           mpm_module             => 'false',
>           name                   => 'Apache',
>           package_ensure         => 'installed',
> 
> It illustrates that the composition layer is not consistent with the
> upstream module data schema, and that could be a potential regression in
> deployment (new parameter added upstream and goes with defaults, being
> ignored by the composition manifest).
> The solution is the same as for the case A.
> 
> [0] https://bugs.launchpad.net/fuel/+bug/1508482
> [1] https://review.openstack.org/240015

Please use the 6th revision to catch the idea. Next revisions are filled
with tons of auto-generated templates representing committed state of
the deployment data plane. This is still WIP...

We are thinking on should we check the data state for *all* classes and
defines of each deployment task, or should we do this only for some of
the classes. The latter one would drastically decrease the amount of
auto-generated data templates. But the former one would allow to catch
more regressions.

> [2]
> https://github.com/openstack/fuel-library/tree/master/tests/noop/astute.yaml
> [3]
> https://review.openstack.org/#/c/240015/7/utils/jenkins/fuel_noop_tests.rb
> 


-- 
Best regards,
Bogdan Dobrelya,
Irc #bogdando