Using Rally on Heat gates¶

Check performance or how to detect regression¶

The easiest way of using Rally is to execute already existing scenarios. One of the examples is presented in patch https://review.openstack.org/#/c/279450/ . In this patch was executed scenario already existing in Rally HeatStacks.create_and_delete_stack. During executing this scenario Rally creates and then, when stack is created, delete Heat stack. All existing scenarios can be found here: https://github.com/openstack/rally-openstack/blob/master/rally_openstack/scenarios/heat/stacks.py

Mentioned scenario uses Heat template as a parameter for task. The template path should be mentioned for argument template_path. It can be one of Heat templates presented in Rally repository (https://github.com/openstack/rally-openstack/tree/master/samples/tasks/scenarios/heat/templates) or new one, like it was done for mentioned patch. New added template should be placed in rally-scenarios/extra/ directory.

Also it’s possible to specify other fields for each Rally task, like sla or context. More information about other configuration setting is available by link https://rally.readthedocs.io/en/latest/plugins/#rally-plugins Mentioned patch was proposed for confirmation caching mechanism of Heat template validation process (see https://specs.openstack.org/openstack/heat-specs/specs/liberty/constraint-validation-cache.html). So it contains some changes in OS::Heat::TestResource resource, which allows to demonstrate mentioned caching feature improvements.

Initially test was run against current devstack installation, where caching is disabled (e.g. Patch Set 7). The follow results were gotten:

In the next patch set (Patch Set 8) was updated by adding Depends-On reference to commit message. It let to execute the same test with patch for devstack, which turns on caching (https://review.openstack.org/#/c/279400/). The results for this case were:

Comparison average values for create_stack action in the first and the second executions shows, that with enabled caching create_stack works faster in 3 times. It is a tangible improvement for create_stack operation. Need to note, that in described test delay for each constraint validation request takes 0.3 sec. as specified in constraint_prop_secs property of TestResource. It may be more, than real time delay, but it allows to confirm, that caching works correct.

Also this approach may be used for detecting regressions. In this case workflow may be presented as follow list of steps:

• add to task list (heat-fakevirt.yaml) existing or new tasks.
• wait a result of this execution.
• upload patchset with changes (new feature) and launch the same test again.
• compare performance results.

Compare output API performance¶

Another example of using Rally job is writing custom Rally scenarios in Heat repository. There is an example of this is presented on review: https://review.openstack.org/#/c/270225/

It’s similar on the first example, but requires more Rally specific coding. New tasks in heat-fakevirt.yaml use undefined in Rally repository scenarios:

• CustomHeatBenchmark.create_stack_and_show_output_new
• CustomHeatBenchmark.create_stack_and_show_output_old
• CustomHeatBenchmark.create_stack_and_list_output_new
• CustomHeatBenchmark.create_stack_and_list_output_old

All these scenarios are defined in the same patch and placed in rally-scenarios/plugins/ directory.

The aim of these scenarios and tasks is to demonstrate differences between new and old API calls. Heat client has a two commands for operating stack outputs: heat output-list and heat output-show <output-id>. Previously there are no special API calls for getting this information from server and this data was obtained from whole Heat Stack object. This was changed after implementation new API for outputs: https://specs.openstack.org/openstack/heat-specs/specs/mitaka/api-calls-for-output.html

As described in the mentioned specification outputs can be obtained via special requests to Heat API. According to this changes code in Heat client was updated to use new API, if it’s available.

The initial problem for this change was performance issue, which can be formulated as: execution command heat output-show <output-id> with old approach required resolving all outputs in Heat Stack, before getting only one output specified by user.

The same issue was and with heat output-list, which required to resolve all outputs only for providing list of output keys without resolved values.

Two scenarios with suffix *_new use new output API. These scenarios are not presented in Rally yet, because it’s new API. Another two scenarios with suffix *_old are based on the old approach of getting outputs. This code was partially replaced by new API, so it’s not possible to use it on fresh devstack. As result this custom code was written as two custom scenarios.

All these scenarios were added to task list and executed in the same time. Results of execution are shown below: