Policy Types¶

Develop A New Policy Type¶

The first step for adding a new policy type is to create a new file containing a subclass of Policy. Then you will define the spec schema for the new policy type in a Python dictionary named spec_schema.

Defining Spec Schema¶

Each key in this dictionary represents a property name; the value of it is an object of one of the schema types listed below:

• String: A string property.
• Boolean: A boolean property.
• Integer: An integer property.
• List: A property containing a list of values.
• Map: A property containing a map of key-value pairs.

For example:

spec_schema = {
  'destroy_after_delete': schema.Boolean(
    'Boolean indicating whether object will be destroyed after deletion.',
    default=True,
  ),
  ...
}

If a property value will be a list, you can further define the type of items the list can accept. For example:

spec_schema = {
  'criteria': schema.List(
    'Criteria for object selection that will be evaluated in order.',
    schema=schema.String('Name of a criteron'),
  ),
  ...
}

If a property value will be a map of key-value pairs, you can define the schema of the map, which is another Python dictionary containing definitions of properties. For example:

spec_schema = {
  'strategy': schema.Map(
    'Strategy for dealing with servers with different states.',
    schema={
      'inactive': 'boot',
      'deleted': 'create',
      'suspended': 'resume',
    },
  ),
  ...
}

When creating a schema type object, you can specify the following keyword arguments to gain a better control of the property:

• default: a default value of the expected data type;
• required: a boolean value indicating whether a missing of the property is acceptable when validating the policy spec;
• constraints: a list of Constraint objects each of which defines a constraint to be checked. Senlin currently only support AllowedValues constraint.

Applicable Profile Types¶

Not all policy types can be used on all profile types. For example, a policy about load-balancing is only meaningful for objects that can handle workloads, or more specifically, objects that expose service access point on an IP port.

You can define what are the profile types your new policy type can handle by specifying the PROFILE_TYPE property of your policy type class. The value of PROFILE_TYPE is a list of profile type names. If a policy type is designed to handle all profile types, you can specify a single entry ANY as the value. See profile types for profile type related operations.

Policy Targets¶

A policy type is usually defined to handle certain operations. The rules embedded in the implementation may need to be checked before the execution of an action or they may need to be enforced after the execution of the action. When an action is about to be executed or an action has finished execution, the Senlin engine will check if any policy objects attached to a cluster is interested in the action. If the answer is yes, the engine will invoke the pre_op function or the post_op function respectively, thus giving the policy object a chance to adjust the action’s behavior.

You can define a TARGET property for the policy type implementation to indicate the actions your policy type want to subscribe to. The TARGET property is a list of tuple (WHEN, ACTION). For example, the following property definition indicates that the policy type is interested in the action CLUSTER_SCALE_IN and CLUSTER_DEL_NODES. The policy type wants itself be consulted before these actions are performed.

class MyPolicyType(Policy):
  ...
  TARGET = [
    (BEFORE, consts.CLUSTER_SCALE_IN),
    (BEFORE, consts.CLUSTER_DEL_NODES),
  ]
  ...

When the corresponding actions are about to be executed, the pre_op function of this policy object will be invoked.

Passing Data Between Policies¶

Each policy type may decide to send some data as additional inputs or constraints for the action to consume. This is done by modifying the data property of an Action object (see action).

A policy type may want to check if there are other policy objects leaving some policy decisions in the data property of an action object.

Senlin allows for more than one policies to be attached to the same cluster. Each policy, when enabled, is supposed to check a specific subset of cluster actions. In other words, different policies may get checked before/after the engine executes a specific cluster action. This design is effectively forming a chain of policies for checking. The decisions (outcomes) from one policy somtimes impact other policies that are checked later.

To help other developers to understand how a specific policy type is designed to work in concert with others, we require all policy type implementations shipped with Senlin accompanied by a documentation about:

• the action data items the policy type will consume, including how these data will impact the policy decisions.
• the action.data items the policy type will produce, thus consumable by any policies downstream.

For built-in policy types, the protocol is documented below:

Registering The New Policy Type¶

For Senlin service to be aware of and thus to make use of the new policy type you have just developed, you will register it to the Senlin service. Currently, this is done through a manual process shown below. In future, Senlin will provide dynamical loading support to policy type plugins.

To register a new plugin type, you will add a line to the setup.cfg file that can be found at the root directory of Senlin code base. For example:

[entry_points]
senlin.policies =
    ScalingPolicy = senlin.policies.scaling_policy:ScalingPolicy
    MyCoolPolicy = <path to the policy module>:<policy class name>

Finally, save that file and do a reinstall of the Senlin service, followed by a restart of the senlin-engine process.

$ sudo pip install -e .

Now, when you do a senlin policy-type-list, you will see your policy type listed along with other existing policy types.