This document describes the ONAP Policy Framework. It lays out the architecture of the framework and specifies the APIs provided to other components that interwork with the framework. It describes the implementation of the framework, mapping out the components, software strucure, and execution ecosystem of the framework. It goes on to provide examples that illustrate how to write, deploy, and run policies of various types using the framework.

1. Overview

The ONAP Policy Framework is a comprehensive policy design, deployment, and execution environment. The Policy Framework is the decision making component in an ONAP system. It allows you to specify, deploy, and execute the governance of the features and functions in your ONAP system, be they closed loop, orchestration, or more traditional open loop use case implementations. The Policy Framework is the component that is the source of truth for all policy decisions.

One of the most important goals of the Policy Framework is to support Policy Driven Operational Management during the execution of ONAP control loops at run time. In addition, use case implementations such as orchestration and control benefit from the ONAP policy Framework because they can use the capabilities of the framework to manage and execute their policies rather than embedding the decision making in their applications.

The Policy Framework is deployment agnostic, the Policy Framework manages Policy Execution (in PDPs) and Enforcement (in PEPs) regardless of how the PDPs and PEPs are deployed. This allows policy execution and enforcement can be deployed in a manner that meets the performance requirements of a given application or use case. In one deployment, policy execution could be deployed in a separate executing entity in a Docker container. In another, policy execution could be co-deployed with an application to increase performance.

The ONAP Policy Framework architecture separates policies from the platform that is supporting them. The framework supports development, deployment, and execution of any type of policy in ONAP. The Policy Framework is metadata (model) driven so that policy development, deployment, and execution is as flexible as possible and can support modern rapid development ways of working such as DevOps. A metadata driven approach also allows the amount of programmed support required for policies to be reduced or ideally eliminated.

We have identified five capabilities as being essential for the framework:

  1. Most obviously, the framework must be capable of being triggered by an event or invoked, and making decisions at run time.
  2. It must be deployment agnostic; capable of managing policies for various Policy Decision Points (PDPs) or policy engines.
  3. It must be metadata driven, allowing policies to be deployed, modified, upgraded, and removed as the system executes.
  4. It must provide a flexible model driven policy design approach for policy type programming and specification of policies.
  5. It must be extensible, allowing straightforward integration of new PDPs, policy formats, and policy development environments.

Another important aim of the architecture of a model driven policy framework is that it enables much more flexible policy specification. The ONAP Policy Framework complies with the TOSCA modelling approach for policies, see the TOSCA Policy Primer for more information on how policies are modelled in TOSCA.

  1. A Policy Type is a general implementation of a policy for a feature. For example, a Policy Type could be written to manage Service Level Agreements for VPNs. The Policy Type is designed by a domain expert, who specifies the parameters, triggers, and actions that the Policy Type will have. The implementation (the logic, rules, and tasks of the Policy Type) is implemented by a skilled policy developer in consultation with domain experts.
    1. For example, the VPN Policy Type is used to create VPN policies for a bank network, a car dealership network, or a university with many campuses.
    2. In ONAP, specific ONAP Policy Types are used to create specific policies that drive the ONAP Platform and Components.
  2. A Policy is created by configuring a Policy Type with parameters. For example, the SLA values in the car dealership VPN policy for a particular dealership are configured with values appropriate for the expected level of activity in that dealership.

The ONAP Policy Framework for building, configuring and deploying PDPs is extendable. It allows the use of ONAP PDPs as is, the extension of ONAP PDPs, and lastly provides the capability for users to create and deploy their own PDPs. The ONAP Policy Framework provides distributed policy management for all policies in ONAP at run time. Not only does this provide unified policy access and version control, it provides life cycle control for policies and allows detection of conflicts across all policies running in an ONAP installation.

2. Architecture

The diagram below shows the architecture of the ONAP Policy Framework at its highest level.

The PolicyDevelopment component implements the functionality for development of policy types and policies. PolicyAdministration is responsible for the deployment life cycle of policies as well as interworking with the mechanisms required to orchestrate the nodes and containers on which policies run. PolicyAdministration is also responsible for the administration of policies at run time; ensuring that policies are available to users, that policies are executing correctly, and that the state and status of policies is monitored. PolicyExecution is the set of PDPs running in the ONAP system and is responsible for making policy decisions and for managing the administrative state of the PDPs as directed by PolicyAdministration. 

PolicyDevelopment creates policy artifacts and supporting information in the policy database. PolicyAdministration reads those artifacts and the supporting information from the policy database whilst deploying policy artifacts. Once the policy artifacts are deployed, PolicyAdministration handles the run-time management of the PDPs on which the policies are running. PolicyDevelopment interacts with ONAP design time components, and has no programmatic interface with PolicyAdministration, PolicyExecution or any other run-time ONAP components.

The diagram below shows a more detailed view of the architecture, as inspired by RFC-2753 and RFC-3198.

PolicyDevelopment provides a CRUD API for policy types and policies. The policy types and policy artifacts and their metadata (Information about policies, policy types, and their interrelations ) are stored in the PolicyDB. The PolicyDevGUI, PolicyDistribution, and other applications such as Portal can use the PolicyDevelopment API to create, update, and delete policy types and policies.

PolicyAdministration has two important functions:

  • Management of the life cycle of PDPs in an ONAP installation. PDPs register with PolicyAdministration when they come up. PolicyAdministration handles the allocation of PDPs to a PDP Groups and PDP Subgroups, so that they can be managed as microservices in Kubernetes.
  • Management of the deployment of policies to PDPs in an ONAP installation. PolicyAdministration gives each PDP group a set of domain policies to execute.

PolicyAdministration handles PDPs and policy allocation to PDPs using asynchronous messaging over DMaaP.

PolicyAdministation provides three APIs:

  • a CRUD API for policy groups and subgroups
  • an API that allows the allocation of policies PDP groups and subgroups to be controlled
  • an API allows policy execution to be managed, showing the status of policy execution on PDP Groups, subgroups, and individual PDPs as well as the life cycle state of PDPs

PolicyExecution is the set of running PDPs that are executing policies, logically partitioned into PDP groups and subgroups.

The figure above shows how PolicyExecution looks at run time with PDPs running in Kubernetes. A PDPGroup is a purely logical construct that collects all the PDPs that are running policies for a particular domain together. A PDPSubGroup is a group of PDPs of the same type that are running the same policies. A PDPSubGroup is deployed as a Kubernetes Deployment. PDPs are defined as Kubernetes Pods. At run time,  the actual number of PDPs in each PDPSubGroup is specified in the configuration of the Deployment of that PDPSubGroup in Kubernetes. This structuring of PDPs is required because, in order to simplify deployment and scaling of PDPs in Kubernetes, we gather all the PDPs of the same type that are running the same policies together for deployment.

For example, assume we have policies for the SON (Self Organizing Network) and ACPE (Advanced Customer Premises Service) domains. For SON, we have XACML, Drools, and APEX policies, and for ACPE we have XACML and Drools policies. The table below shows the resultingPDPGroup, PDPSubGroup, and PDP allocations:

PDP GroupPDP SubgroupKubernetes DeploymentKubernetes Deployment StrategyPDPs in Pods


SON-XACML-DepAlways 2, be geo redundant2 PDP-X


SON-Drools-DepAt Least 4, scale up on 70% load, scale down on 40% load, be geo-redundant>= 4 PDP-D
SON-APEXSON-APEX-DepAt Least 3, scale up on 70% load, scale down on 40% load, be geo-redundant>= 3 PDP-A


ACPE-DroolsACPE-Drools-DepAt Least 2, scale up on 80% load, scale down on 50% load>=2 PDP-D

2.1 Policy Framework Object Model

This section describes the structure of and relations between the main concepts in the Policy Framework. This model is implemented as a common model and is used by PolicyDevelopment, PolicyDeployment, and PolicyExecution.

The UML class diagram above shows the portion of the Policy Framework Object Model that applies to PolicyDeployment and PolicyExecution.

The UML class diagram above shows the portion of the Policy Framework Object Model that applies to PolicyDevelopment and PolicyDeployment.

2.2 Policy Design Architecture

This section describes the architecture of the model driven system used to develop policy types and to create concrete policies using policy types. The output of Policy Design is deployment-ready artifacts and Policy metadata in the Policy Framework database.

Policies that are expressed via natural language or a model require some development work ahead of time for them to be translated into concrete runtime policies. Some Policy Domains will be setup and available in the platform during startup such as Control Loop Operational Policy Models, OOF placement Models, DCAE microservice models. Policy type implementation development is done by an experienced developer.

2.2.1 Policy Type Design

Policy Type Design is the task of creating policy types that capture the generic and vendor independent aspects of a policy for a particular domain use case. The policy type implementation specifies the model information, rules, and tasks that a policy type requires to generate concrete policies.

All policy types must implement the ONAP Policy Framework PolicyType interface. This interface allows PolicyDevelopment to manage policy types and to generate policies from these policy types in a uniform way regardless of the domain that the policy type is addressing or the PDP technology that will execute the policy. The interface is used by PolicyDevelopment to determine the PDP technology of the policy type, the structure, type, and definition of the model information that must be supplied to the policy type to generate a concrete policy.

PolicyTypeImpl is developed for a certain type of PDP (for example XACML oriented for decision policies or Drools rules oriented for ECA policies). The design environment and tool chain for a policy type is specific for the type of policy being designed.

The PolicyTypeImpl  implementation (or raw policy) is the specification of the specific rules or tasks, the flow of the policy, its internal states and data structures and other relevant information. A PolicyTypeImpl is specific to a PDP technology, that is XACML, Drools, or APEX. A PolicyTypeImpl can be specific to a particular policy type, it can be more general, providing the implementation of a class of policy types, or the same policy type may have many implementations.

PolicyDevelopment provides the RESTful Policy Design API, which allows other components to query policy types and policy type implementations, to determine the model information, rules, or tasks that they require, to specialize policy flow, and to generate policies from policy types. This API is used by the ONAP Policy Framework and other components such as PolicyDistribution to create policies from policy types.

Consider a policy type created for managing faults on vCPE equipment in a vendor independent way. The policy type captures the generic logic required to manage the faults and specifies the vendor specific information that must be supplied to the type for specific vendor vCPE VFs. The actual  vCPE policy that is used for managing particular vCPE equipment is created by setting the parameters specified in the policy type together with the specific modeled information, rules and tasks in the policy type implementation for that vendor model of vCPE.

2.2.1 Generating Policy Types

It is possible to generate policy types using MDD (Model Driven Development) techniques. Policy types are expressed using a DSL (Domain Specific Language) or a policy specification environment for a particular application domain. For example, policy types for specifying SLAs could be expressed in a SLA DSL and policy types for managing SON features could be generated from a visual SON management tool. The ONAP Policy framework provides an API that allows tool chains to create policy types. SDC uses this approach for generating Policy Types in the Policy Framework, see the Model driven Control Loop Design page.

The SDC GUI supports several types of policies that can be captured at design time. DCAE micro service configuration policies can be onboarded via the DCAE-DS (DCAE Design Studio).

Policy Type Design with SDC DCAE-DS PolicyDesign «API_User»DCAE_DS «API_User»DCAE_DS PolicyTypeDesign PolicyTypeDesign PolicyDB PolicyDB 1Get Policy Type Reference and Metadata 2Get Policy Type Metadataand Artifact alt[Policy Type Artifact exists] 3Return Policy Type Reference and Metadata [Policy Artifact does not exist] 4Return New Policy Type Reference and Empty Metadata 5Policy Type Editing and Generation Session 6Create Policy Type and Metadata 7Policy Generation Result

The GUI implementation in another ONAP component such as SDC DCAE-DS uses the API_User API to create and edit ONAP policy types. Programming Policy Type Implementations

For skilled developers, the most straightforward way to create a policy type is to program it. Programming a policy type might simply mean creating and editing text files, thus manually creating the TOSCA Policy Type Yaml file and the policy type implementation for the policy type.

A more formal approach is preferred. For policy type implementations, programmers use a specific Eclipse project type for developing each type of implementation, a Policy Type Implementation SDK. The project is under source control in git. This Eclipse project is structured correctly for creating implementations for a specific type of PDP. It includes the correct POM files for generating the policy type implementation and has editors and perspectives that aid  programmers in their work

2.2.2 Policy Design

The PolicyCreation function of PolicyDevelopment creates policies from a policy type.  The information expressed during policy type design is used to parameterize a policy type to create an executable policy. A service designer and/or operations team can use tooling that reads the TOSCA Policy Type specifications to express and capture a policy at its highest abstraction level. Alternatively, the parameter for the policy can be expressed in a raw JSON or YAML file and posted over the policy design API described on the Model driven Control Loop Design page.

A number of mechanisms for  policy creation are supported in ONAP. The process in PolicyDevelopment for creating a policy is the same for all mechanisms. The most general mechanism for creating a policy is using the RESTful Policy Design API, which provides a full interface to the policy creation support of PolicyDevelopment. This API may be exercised directly using utilities such as curlPolicyDevelopment provides a command line tool that is a loose wrapper around the API. It also provides a general purpose Policy GUI in the ONAP Portal for policy creation, which again is a general purpose wrapper around the policy creation API. The Policy GUI can interpret any TOSCA Model ingested and flexibly presents a GUI for a user to create policies from. The development of these mechanisms will be phased over a number of ONAP releases.

A number of ONAP components use policy in manners which are specific to their particular needs. The manner in which the policy creation process is triggered and the way in which information required to create a policy is specified and accessed is specialized for these ONAP components.

The following subsections outline the mechanisms for policy creation and modification supported by the ONAP Policy Framework. Policy Design in the ONAP Policy Framework

Policy creation in PolicyDevelopment follows the general sequence shown in the sequence diagram below. An API_USER is any component that wants to create a policy from a policy type. PolicyDevelopment supplies a REST interface that exposes the API and also provides a command line tool and general purpose client that wraps the API.

Policy Design over the REST API PolicyDesign API_User API_User PolicyDesign PolicyDesignPolicyDB PolicyDB 1Get Policy Type Reference 2Get Policy Type Artifact and Metadata 3Return Policy Type Reference and Metadata 4Get Policy Reference and Metadata 5Get Policy Metadata alt[Policy Artifact exists] 6Return Policy Reference and Metadata [Policy Artifact does not exist] 7Return New Policy Reference and Empty Metadata 8Policy Editing and Generation Sessionto get Policy Parameters from user 9Use Policy Type specification 10Create Policy 11Create Policy 12Save Policy Artifact and Metadata 13Policy Creation Result

A PolicyDevAPIUser first gets a reference to and the metadata for the Policy type for the policy they want to work on from PolicyDevelopmentPolicyDevelopment reads the metadata and artifact for the policy type from the database. The API_User then asks for a reference and the metadata for the policy. PolicyDevelopment looks up the policy in the database. If the policy already exists, PolicyDevelopment reads the artifact and returns the reference of the existing policy to the PolicyDevAPIUser with the metadata for the existing policy. If the policy does not exist, PolicyDevelopment creates and new reference and metadata and returns that to the API_User.

The PolicyDevAPIUser may now proceed with a policy specification session, where the parameters are set for the policy using the policy type specification. Once the PolicyDevAPIUser is happy that the policy is completely and correctly specified, it requests PolicyDevelopment to create the policy. PolicyDevelopment creates the policy, stores the created policy artifact and its metadata in the database. Model Driven VF (Virtual Function) Policy Design via VNF SDK Packaging

VF vendors express policies such as SLA, Licenses, hardware placement, run-time metric suggestions, etc. These details are captured within the VNF SDK and uploaded into the SDC Catalog. The SDC Distribution APIs are used to interact with SDC. For example, SLA and placement policies may be captured via TOSCA specification. License policies can be captured via TOSCA or an XACML specification. Run-time metric vendor recommendations can be captured via VES Standard specification.

The sequence diagram below is a high level view of SDC-triggered concrete policy generation for some arbitrary entity EntityA. The parameters to create a policy are read from a TOSCA Policy specification read from a CSAR received from SDC.

Model Driven Policy Design PolicyDesign SDC SDC DMaaP DMaaP «API_User»PolicyDistribution «API_User»PolicyDistribution PolicyDesign PolicyDesign 1Subscribe for SDC Notificationson Entries of Interest 2New Entry of Interest: EntityA 3New Entry of Interest: EntityA 4Get Policy Reference 5Return Policy Reference and Metadata 6Create Policy Editing and Generation Session loop 7Use Policy Type specification toread Policy Parametersfrom TOSCA specification in CSAR 8Create Policy 9Policy Creation Result

PolicyDesign uses the PolicyDistribution component for managing SDC-triggered  policy creation and update requests. PolicyDistribution is an API_User, it uses the Policy Design API for policy creation and update. It reads the information it needs to populate the policy type from a TOSCA specification in a CSAR received from SDC and then uses this information to automatically generate a policy.

Note that SDC provides a wrapper for the SDC API as a Java Client and also provides a TOSCA parser. See Policy Platform - SDC Service Distribution Software Architecture

In Step 4 above, the PolicyDesign must download the CSAR file. If the policy is to be composed from the TOSCA definition, it must also parse the TOSCA definition.

In Step 9 above, the PolicyDesign must send back/publish status events to SDC such as DOWNLOAD_OK, DOWNLOAD_ERROR, DEPLOY_OK, DEPLOY_ERROR, NOTIFIED. Scripted Model Driven Policy Design

Service policies such as optimization and placement policies can be specified as a TOSCA Policy at design time. These policies use a TOSCA Policy Type specification as their schemas. Therefore, scripts can be used to create TOSCA policies using TOSCA Policy Types.

Scripted Policy Design PolicyDesign «API_User»Script «API_User»Script DirectiveFile DirectiveFile PolicyTypeDesign PolicyTypeDesign PolicyDesign PolicyDesign ConcretePolicyDesign ConcretePolicyDesign loop 1Read next directive from script file 2Read Policy Type for directive 3Create TOSCA Policy 4Read parameters from script file directive 5Set Parameters in TOSCA Policy 6Generate Policy 7Policy Generation Resultend 8Generate Policy 9Policy Generation Result

One straightforward way of generating policies from Policy types is to use directives specified in a script file. The command line utility is an API_User. The script reads directives from a file. For each directive, it reads the policy type using the Policy Type API, and uses the parameters of the directive to create a TOSCA Policy. It then uses the Policy API to create the policy.

2.2.3 Policy Design Process

All policy types must be certified as being fit for deployment prior to run time deployment. In the case of design-time via the SDC application, it is assumed the lifecycle being implemented by SDC will suffice for any policy types that are declared within the ONAP Service CSAR. For other policy types and policy type implementations, the lifecycle associated with software development process will suffice. Since policy types and their implementations will be designed and implemented using software development best practices, they can be utilized and configured for various environments (eg. development, testing, production) as desired.

2.3 Policy Runtime Architecture

The Policy Framework Platform components are themselves designed as micro services that are easy to configure and deploy via Docker images and K8S both supporting resiliency and scalability if required. PAPs and PDPs are deployed by the underlying ONAP management infrastructure and are designed to comply with the ONAP interfaces for deploying containers.

The PAPs keep track of PDPs, support the deployment of PDP groups and the deployment of a policy set across those PDP groups. A PAP is stateless in a RESTful sense. Therefore, if there is more than one PAP deployed, it does not matter which PAP a user contacts to handle a request. The PAP uses the database (persistent storage) to keep track of ongoing sessions with clients. Policy management on PDPs is the responsibility of PAPs; management of policy sets or policies by any other manner is not permitted.

In the ONAP Policy Framework, the interfaces to the PDP are designed to be as streamlined as possible. Because the PDP is the main unit of scalability in the Policy Framework, the PF is designed to allow PDPs in a PDP group to arbitrarily appear and disappear and for policy consistency across all PDPs in a PDP group to be easily maintained. Therefore, PDPs have just two interfaces; an interface that users can use to execute policies and interface to the PAP for administration, life cycle management and monitoring. The PAP is responsible for controlling the state across the PDPs in a PDP group. The PAP interacts with the Policy database and transfers policy sets to PDPs, and may cache the policy sets for PDP groups.

See also Sectino 2 of the TO BE DELETED - refer to Dublin Documentation page, where the mechanisms for PDP Deployment and Registration with PAP are explained.

2.3.1 Policy Framework Services

The ONAP Policy Framework follows the architectural approach for micro services recommended by the ONAP Architecture Subcommittee.

The ONAP Policy Framework defines Kubernetes Services to manage the life cycle of Policy Framework executable components at runtime. A Kubernetes service allows, among other parameters,  the number of instances (pods in Kubernetes terminology) that should be deployed for a particular service to be specified and a common endpoint for that service to be defined. Once the service is started in Kubernetes, Kubernetes ensures that the specified number of instances is always kept running. As requests are received on the common endpoint, they are distributed across the service instances. More complex call distribution and instance deployment strategies may be used; please see the Kubernetes Services documentation for those details.

If, for example, a service called policy-pdpd-control-loop is defined that runs 5 PDP-D instances. The service has the end point https://policy-pdpd-control-loop.onap/<service-specific-path>. When the service is started, Kubernetes spins up 5 PDP-Ds. Calls to the end point https://policy-pdpd-control-loop.onap/<service-specific-path> are distributed across the 5 PDP-D instances. Note that the .onap part of the service endpoint is the namespace being used and is specified for the full ONAP Kuberentes installation.

The following services will be required for the ONAP Policy Framework:

PAPhttps://policy-papThe PAP service, used for policy administration and deployment. See TO BE DELETED - refer to Dublin Documentation for details of the API for this service

A PDP service is defined for each PDP group. A PDP group is identified by the domain on which it operates.

For example, there could be two PDP-X domains, one for admission policies for ONAP proper and another for admission policies for VNFs of operator Supacom. Two PDP-X services are defined:



There is one and only one PAP service, which handles policy deployment, administration, and monitoring for all policies in all PDPs and PDP groups in the system. There are multiple PDP services, one PDP service for each domain for which there are policies.

2.3.2 The Policy Framework Information Structure

The following diagram captures the relationship between Policy Framework concepts at run time.

Runtime Relationships between Concepts PDPSubGroup PDPService PolicySet PDP PolicyImpl LifecycleManagedBy11 ManagesLifecycleOf1* Executes11 ExecutesOn1* Contains1*

There is a one to one relationship between a PDP SubGroup, a Kubernetes PDP service, and the set of policies assigned to run in the PDP subgroup. Each PDP service runs a single PDP subgroup with multiple PDPs, which executes a specific Policy Set containing a number of policies that have been assigned to that PDP subgroup. Having and maintaining this principle makes policy deployment and administration much more straightforward than it would be if complex relationships between PDP services, PDP subgroups, and policy sets.

The topology of the PDPs and their policy sets is held in the Policy Framework database and is administered by the PAP service.


The diagram above gives an indicative structure of the run time topology information in the Policy Framework database. Note that the PDP_SUBGROUP_STATE and PDP_STATE fields hold state information for life cycle management of PDP groups and PDPs.

2.3.3 Startup, Shutdown and Restart

This section describes the interactions between Policy Framework components themselves and with other ONAP components at startup, shutdown and restart. PAP Startup and Shutdown

The sequence diagram below shows the actions of the PAP at startup.

PAP Startup and Shutdown K8S K8S PAP PAPPolicyDB PolicyDB 1Start PAP 2Read Initial Data 3Initialize Policy Administration loop[forever] refWait for andHandle PDPStatus Updates refWait for andHandleAdmin Requests

The PAP is the run time point of coordination for the ONAP Policy Framework. When it is started, it initializes itself using data from the database.  It then waits for periodic PDP status updates and for administration requests.

PAP shutdown is trivial. On receipt or a shutdown request, the PAP completes or aborts any ongoing operations and shuts down gracefully. PDP Startup and Shutdown

The sequence diagram below shows the actions of the PDP at startup. See also Section 4 of the TO BE DELETED - refer to Dublin Documentation page for the API used to implement this sequence.

PDP Startup and Shutdown K8S K8S PDP PDP DMaaP DMaaP PAP PAP 1Start PDP 2Send Status (initial status) to PAP refPAP downloads Policy Set to PDP loop[forever] 3Send Status to PAP

At startup, the PDP initializes itself.  At this point it is in PASSIVE mode. The PDP begins sending periodic Status messages to the PAP.

The first Status message initializes the process of loading the correct Policy Set on the PDP in the PAP.

On receipt or a shutdown request, the PDP completes or aborts any ongoing policy executions and shuts down gracefully.

2.3.4 Policy Execution

Policy execution is the execution of a policy in a PDP. Policy enforcement occurs in the component that receives a policy decision.

Policy Execution REQUESTOR REQUESTOR RESPONDEE RESPONDEE PDP PDP alt[Synchronous Invocation] 1Execute Policy wait 2Policy Execution Result [Asynchronous Invocation] 3Execute Policy 4Policy Execution Result

Policy execution can be synchronous or asynchronous. In synchronous policy execution, the component requesting a policy decision requests a policy decision and waits for the result. The PDP-X and PDP-A use synchronous policy execution. In asynchronous policy execution, the component that requests a policy decision does not wait for the decision. Indeed, the decision may be passed to another component. The PDP-D and PDP-A use asynchronous policy execution.

Policy execution is carried out using the current life cycle mode of operation of the PDP. While the actual implementation of the mode may vary somewhat between PDPs of different types, the principles below hold true for all PDP types:

Lifecycle ModeBehaviour
PASSIVE MODEPolicy execution is always rejected irrespective of PDP type.
ACTIVE MODEPolicy execution is executed in the live environment by the PDP.
SAFE MODEPolicy execution proceeds, but changes to domain state or context are not carried out. The PDP returns an indication that it is running in SAFE mode together with the action it would have performed if it was operating in ACTIVE mode. The PDP type and the policy types it is running must support SAFE mode operation.
TEST MODEPolicy execution proceeds and changes to domain and state are carried out in a test or sandbox environment. The PDP returns an indication it is running in TEST mode together with the action it has performed on the test environment. The PDP type and the policy types it is running must support TEST mode operation.

2.3.5 Policy Lifecycle Management

Policy lifecycle management manages the deployment and life cycle of policies in PDP groups at run time. Policy sets can be deploy at run time without restarting PDPs or stopping policy execution. PDPs preserve state for minor/patch version upgrades and rollbacks. Load/Update Policies on PDP

The sequence diagram below shows how policies are loaded or updated on a PDP.

Download Policy Set to PDP User User PolicyDB PolicyDB PAP PAP DMaaP DMaaP PDP PDP alt[PDP Startup] 1Send Status (initial status) to PAP [Update Request from User] 2Update Policy on PDP 3Read PDP metadata and Policy Set 4Prepare for PDP update 5Set PDP mode as PASSIVE 6Transfer Policy Set Policies to PDP 7Load Policies alt[PDP Mode should be ACTIVE] 8Set PDP mode as ACTIVE refExecute Policies inACTIVE Mode [PDP Mode should be SAFE] 9Set PDP mode as SAFE refExecute Policies inSAFE Mode [PDP Mode should be TEST] 10Set PDP mode as TEST refExecute Policies inTEST Mode 11Update PDP metadataThis sequence can be initiated in two ways; from the PDP or from a user action.

  1. A PDP sends regular status update messages to the PAP. If this message indicates that the PDP has no policies or outdated policies loaded, then this sequence is initiated
  2. A user may explicitly trigger this sequence to load policies on a PDP

The PAP controls the entire process. The PAP reads the current PDP metadata and the required policy and policy set artifacts from the database. It then builds the policy set for the PDP. Once the policies are ready, the PAP sets the mode of the PDP to PASSIVE. The Policy Set is transparently passed to the PDP by the PAP. The PDP loads all the policies in the policy set including any models, rules, tasks, or flows in the policy set in the policy implementations.

Once the Policy Set is loaded, the PAP orders the PDP to enter the life cycle mode that has been specified for it (ACTIVE/SAFE/TEST). The PDP beings to execute policies in the specified mode (see section 2.3.4). Policy Rollout

A policy set steps through a number of life cycle modes when it is rolled out.

Policy Rollout User User PAP PAP PDP PDP 1Create Policy Set for PDP Group loop[over PDP Group] refPAP downloads Policy Set to PDP 2Set PDP Group in Test Mode loop[over PDP Group] 3Set PDP in TEST mode 4Report test results 5Report consolidated test results 6Set PDP Group in Safe Mode loop[over PDP Group] 7Set PDP in SAFE mode 8Report safe mode operation results 9Report consolidated safe mode operation results 10Set PDP Group in Active Mode loop[over PDP Group] 11Set PDP in Active mode loop[forever] 12Report active mode operation results 13Report consolidated active mode operation results

The user defines the set of policies for a PDP group. It is deployed to a PDP group and is initially in PASSIVE mode. The user sets the PDP Group into TEST mode. The policies are run in a test or sandboxed environment for a period of time. The test results are passed back to the user. The user may revert the policy set to PASSIVE mode a number of times and upgrade the policy set during test operation.

When the user is satisfied with policy set execution and when quality criteria have been reached for the policy set, the PDP group is set to run in SAFE mode. In this mode, the policies run on the actual target environment but do not actually exercise any actions or change any context in the target environment. Again, as in TEST mode, the operator may decide to revert back to TEST mode or even PASSIVE mode if issues arise with a policy set.

Finally, when the user is satisfied with policy set execution and when quality criteria have been reached, the PDP group is set into ACTIVE state and the policy set executes on the target environment. The results of target operation are reported. The PDP group can be reverted to SAFE, TEST, or even PASSIVE mode at any time if problems arise. Policy Upgrade and Rollback

There are a number of approaches for managing policy upgrade and rollback.

The most straightforward approach is to use the approach described in section for upgrading and rolling back policy sets. In order to upgrade a policy set, one follows the process in with the new policy set version. For rollback, one follows the process in section with the older policy set, most probably setting the old policy set into ACTIVE mode immediately. The advantage of this approach is that the approach is straightforward. The obvious disadvantage is that the PDP group is not executing on the target environment while the new policy set is in PASSIVE, TEST, and SAFE mode. 

A second manner to tackle upgrade and rollback is to use a spare-wheel approach. An special upgrade PDP group service is set up as a K8S service in parallel with the active one during the upgrade procedure. The spare wheel service is used to execute the process described in section When the time comes to activate the policy set, the references for the active and spare wheel services are simply swapped. The advantage of this approach is that the down time during upgrade is minimized, the spare wheel PDP group can be abandoned at any time without affecting the in service PDP group, and the upgrade can be rolled back easily for a period simply by preserving the old service for a time. The disadvantage is that this approach is more complex than the first approach.

A third approach is to have two policy sets running in each PDP, an active set and a standby set. However such an approach would increase the complexity of implementation in PDPs significantly.

2.3.6 Policy Monitoring

PDPs provide a periodic report of their status to the PAP. All PDPs report using a standard reporting format that is extended to provide information for specific PDP types. PDPs provide at least the information below:

TimestampTime the report record was generated
InvocationCountThe number of execution invocations the PDP has processed since the last report
LastInvocationTimeThe time taken to process the last execution invocation
AverageInvocationTimeThe average time taken to process an invocation since the last report
StartTimeThe start time of the PDP
UpTimeThe length of time the PDP has been executing
RealTimeInfoReal time information on running policies.

2.3.7 PEP Registration and Enforcement Guidelines

In ONAP there are several applications outside the Policy Framework that enforce policy decisions based on models provided to the Policy Framework. These applications are considered Policy Enforcement Engines (PEP) and roles will be provided to those applications using AAF/CADI to ensure only those applications can make calls to the Policy Decision API's. Some example PEP's are: DCAE, OOF, and SDNC.

See Section 3.4 of the TO BE DELETED - refer to Dublin Documentation for more information on the Decision APIs.

3. APIs Provided by the Policy Framework

See the TO BE DELETED - refer to Dublin Documentation page.

4. Terminology

PAP (Policy Administration Point)A component that administers and manages policies
PDP (Policy Deployment Point)A component that executes a policy artifact (One or many?)
PDP_<>A specific type of PDP
PDP GroupA group of PDPs that execute the same set of policies
Policy DevelopmentThe development environment for policies
Policy Type

A generic prototype definition of a type of policy in TOSCA, see the TOSCA Policy Primer

PolicyAn executable policy defined in TOSCA and created using a Policy Type, see  the TOSCA Policy Primer
Policy SetA set of policies that are deployed on a PDP group. One and only one Policy Set is deployed on a PDP group
  • No labels


  1. Pamela Dragosh I think we should have a discussion/workshop on Policy Design. I'm not sure whether some of the cases above (Reading from SDC and the DCAE case) are better specified as specialization of proto policies into executable policies are are parameterization and instantiation of executable policies.

    1. Yes we need to really flush this out in a working session. I think you and I working from different perspectives right now, both valid. How to tie these design time policies with runtime...

    2. Wondering if we need to separate models and then how design builds on top of the models...

  2. Some comments:

    • Section 2:  Policy execution shouldn't talk to PAP.  PAP should be design time only.
    • Section 2:  PDP's can and do access the DB as well.
    • Section 2:  PDP's may access Nexus, e.g. Drools.
    • Section 2.1:  Does this mean users will be designing policies on the fly in Prod?  No testing environment?  Seems dangerous.
    1. Although the PDPs do access the DB and Nexus today, I'm not sure that they should. It really complicates the admin and management of policies. If we can eliminate those interfaces, then the PDPs really become much simpler to deploy, administer, and scale.

      Would it not be much better if the PAP acted as a run-time mediator, handling all access to the DB and Nexus for the PDPs? This could potentially simplify the architecture a lot without degradation in performance.

      Policies are not designed in production. They are designed, tested (in some sort of sand box), and certified. Everything in Section 2.1 is design time. We'll make that explicit in the text.

      1. Yes I agree. The PDP's should access nexus only - not the DB. The DB is used to organize PDP groups, the policies that are deployed to those groups, and the state for the policies. Probably also configuration information for PDP's. Simplify, simplify, simplify. Keep the PAP simple, and PDP's even simpler.

        The Policy Distribution/Deployment API passes the information to the PDP's.

        We are proposing environments for Policy Template Design that definitely include testing. But I am getting requirements from Operations team that want to stage policies in production and be able to dynamically filter, change policy during runtime. We need to ensure the platform can react very quickly and give them the control they are asking for.

  3. Very nice documentation although it is under construction. I will be working to fill out Sec.5, putting in details of policy creation/deployment/running for real world use cases (most are for RAN domain) that I have been working on as well as current problems for new use cases support.

    Below are my thoughts and ideas that might also help improve policy framework:

    New APIs to send event/request into decision engine like drools and xacml.

    New APIs to CRUD models, such as config policy models, operational policy models, event models, actor models (including actor specific request/response, etc.).

    Do we need to consider a policy design studio which incorporate all designer needs in a central place. It can be integrated with SDC or used standalone. It could be seen as wrapped PAP as well, including lifecycle management, manual CRUD of model/template/policy on a per-domain basis, auto-generation of policies from SDC distributed artifacts, etc.

    Hybrid policy deployment from PAP to PDP/PDP group - manual && auto, for example, to have an intelligent module in PAP for policy auto-deployment - intelligently decide which PDP/PDP group the policy should go to based on policy attributes/metadata, PDP/PDP group attributes/metadata, PDP load balancing status, historical PDP usage/query records (machine learning based prediction can be leveraged), etc.

    1. Thanks for working on section 5!

      On APIs to send event/request: Yes this is really important. We plan to cover that in Section 2.2, and there's the beginnings of the API in Section 3.3 already.

      on APIs for CRUD: Again the plan is to have the PDP/PAP APIs and APIs to the PAPs for CRUD in place. The entire framework should be model driven.  We have to work through it together to get the type and granularity of those models right.

      You are correct on the design studio. I haven't thought too much on how this will look or how it will be built yet. Up to now we have been focusing on getting the APIs that such a design studio would use correct in Section 2.2.

      All the deployment modes you describe in the last paragraph are needed. If we get our deployment mechanisms correct then all of the modes of operation you mention should be rather straightforward to achieve. We cold probably shoot for manual and a rather primitive auto deployment in Casablanca and in later releases look at ML/predictive/ and even policy driven (Recursive!!!) deployment in future releases. Something that might be interesting is which policies should be deployed centrally/at the edge/ in the RAN with maybe lightweight, highly used, but rather dumb policies in the RAN and edge and the more heavyweight policies running centrally.........

      1. Yeah. For the last point you raise, maybe we can also borrow some ideas from data storage/caching strategy in hierarchical data center network in which what data and which level of cache the data should be stored are decided in order to achieve low-latency response and high running performance. We can basically replace their "data" by "policies" and do some extension to the deployment algorithm. But anyways, this is long-term stuff.

  4. Liam Fallon - we have been using Policy Translation instead of Concrete Policy. I'm wondering if we merge the two terminologies.

    1. What we are trying to name is the policy that is created from the policy template. I don't know if either "Concrete Policy" or "Translated Policy" really captures that. "Policy Instance" is not good either.

  5. Liam Fallon - this section is a little unclear to me as to what is expected.

    The following services will be required for the ONAP Policy Framework:

    PAPhttps://policy-papThe PAP service, used for policy administration and deployment. See Section 3.2 for details of the API for this service

    A PDP service is defined for each PDP group. A PDP group is identified by the domain on which it operates.

    For example, there could be two PDP-X domains, one for admission policies for ONAP proper and another for admission policies for VNFs of operator Supacom. Two PDP-X services are defined:



    One functionality we do need is how to tell the clients which endpoint to go to for policy decisions. Especially if we starting putting sets of policies into a group of PDPs. But right now it seems only the pdpx is returning decisions.

    1. I was just using the PDP-X as an example. Any domain implemented on any of the PDP types could return a decision.

      Let's assume for the sake of argument we have four domains as follows:
      optical with PDP-D
      WLan with PDP-A
      Ethernet with PDP-X
      DSL with PDP-A

      We then would have four PDP groups and four services:

      In this way we can address and manage the PDP groups for each domain separately.

  6. Liam Fallon

                  some comments about about SDC integration:

    1. The PolicyDesignSDCHandler needs to do additional two steps, between 3 to 4
      1. Download the csar file from SDC by REST FUL API
      2. Tosca parse it and then decide to Get Policy Reference(like specified feature HPA)
    2. After 9, the PolicyDeisgnSDCHandler needs send back/publish status event like (DOWNLOAD_OK, DOWNLOAD_ERROR, DEPOLY_OK, DEPOLY_ERROR, NOTIFIED) On SDC-DISTR-STATUS-TOPICS

    1. Great! I'll update the flow to show this.

  7. Liam Fallon Pamela Dragosh, I don't fully agree upon this statement "A PDP service is defined for each PDP group. A PDP group is identified by the domain on which it operates." I agree the second half that PDP group could be created/identified by the domain but not the first half. If I understand correctly, PDP services we are referring to are PDP-X, PDP-D, PDP-A, etc. I would believe PDP group should be decoupled from PDP service. PDP group could be defined per domain, application/use case, or geographical deployment like central/edge cloud, in which a mix of PDP services could be co-existing. For instance, for Energy-Saving RAN use case, we could create a PDP group including PDP-X and PDP-D deployed in edge cloud serving NYC market. Actually, for most of real-world control loop use cases, we would use a mix of PDP services since PDP-X is always being used for guard decisions. If we couple PDP service with group, for each use case, we'll have to create multiple groups, especially when we want to involve PDP-X/D/A all together. It is a bit cumbersome to users and hard to manage use cases vs. groups. Do you think it is a good idea to leave the flexibility to users and let them create PDP group and decide what PDP-instances/services to put in ?      

    1. Yes, you make a good point. I thought it might be cleaner to confine a PDP service to a particular PDP type just to make the implementation a bit easier. However we may need another layer of indirection between PDP services (Propabably equivalent to a PDP group) and some concept that groups PDPs of a particular type running a particular set of policies.

  8. It would be great to be able to trace (or even debug!) policy executions, track events in/out, decisions, etc.

    Maybe this could achieved by exploiting policy execution monitoring information?

    It would be great to be able to identify the source of policy triggers, and identify who receives outputs. By analysing this information we might be able to deduce information about system stability? 

    1. John, we do have these messages available now and I intend to carry them over into this software architecture. We just need to expand that model. also need to ensure anyone who writes a template conforms to that messaging structure that we currently support.

      For Control Loops:


      For other policies, we can only do so much on our end. Because policy renders decisions, its up to the other ONAP components to enforce them. So perhaps the logging team can impose requirements for those components to log how they enforced a policy decision so that data is available to machine learning or any analytics, auditing applications.

  9. Is a "Concrete Policy" immutable?   It seems it should be if we are to support rollback operations, and the version being part of the identity.    This may also mean that each "Update" operation results in a new Policy which bumps the version of the policy, so the the "U" in the CRUD has to be further elaborated.  

  10. I think we are saying PDP Groups can have members of heterogenous PDP types based on the discussion above in the comments between Liam Fallonand Chenfei Gao.   This would also mean that the deployed Policy Sets are composed of heterogenous policy types, and PDP Group members should know which ones each can consume (which might be is trivial).   In any case, one consideration is the group to be partitioned by PDP types, which in turn has their corresponding PDP members.

    In terms of communication, it seems that, when addressing PDP Groups as a single unit, in other words multicast communication, DMaaP, may be a better fit, than REST based.   

    1. I have been thinking on this and discussing with Ram Krishna Verma. We have to be very careful to get this right because these are the most fundamental concepts in the system for scaling and if we get this wrong now we will have a big job to change it afterwards.

      I think Jorge is correct, we need to be able to scale by PDP type.

      Consider the case where we have Policy managing the containers going through large ports such as Rotterdam, Shanghai, and Long Beach and we have an Executable Policy Set with PDP-X, PDP-D, and PDP-A policies. In each port, we must be able to dimension the number of PDP-s of each type to suit the local conditions. It might be 2/4/2 in one port or 3/5/4 in another.

      Therefore we need two concepts that allows us to have a concept that allows us to partition the policies in the Executable Policy Set by PDP type and also the PDPs in the PDP group by PDP type. Then we can manage and scale any installation/domain as we wish.

      I'll come back with a suggestion  and maybe we can exchange some thoughts on this in the next few days.

      Jorge Hernandez Pamela Dragosh Chenfei Gao

  11. Liam Fallon Pamela Dragosh Jorge Hernandez Chenfei Gao

    As per my understanding so far, there will be following containers in run time:

    API Engine - To provide life cycle API's (CRUD) of policy, policy-template & event-schema. Also, this component is responsible for uploading policy artifacts to nexus & creating policy metadata in DB.

    PAP Engine - To provide API's for deploying policy, managing PDP Groups, sharing run time status of PDP's. This component is responsible for downloading the policy artifacts from nexus and reading metadata from DB to identify the PDP Group in which the downloaded artifacts has to be deployed. Also, when PDP Engine comes up it will contact this component to register itself and then download all the required policies.

    PDP Engine - To execute the policies provided by PAP Engine. They can be either PDP-X, PDP-D, PDP-A or any other PDP in future.

    Distribution Engine - To listen to notifications from SDC and create policies using the API Engine. Once policy is created, deploy them to a PDP Group using PAP Engine.

    If the above understanding is correct, then all the Policy Design API's under section 3.1 should use API Engine instead of PAP. Is that correct??

    1. I wonder should the distribution engine a first order component or should we move its functionality into API Engine or PAP or parts of it into both?

      1. I think it should be a microservice on its own right, but it seems it loses a little relevance as its Casablanca direct interfaces to PDPs are replaced with a single one to the PAP).   For clarity, may be in Ram Krishna Verma 's description of the Distribution Engine references to the "PDP Group" should be avoided so there's no confusion.

    2. The PAP engine needs some additional functionality for setting Lifecycle state - for example "Safe Mode". We can omit "Test Mode" because we don't understand that - but we need to be able to communicate that for a policy(s) and/or group of PDP(s). We may not have to test that this release, but I would like that functionality in there.

  12. Adding PDP's to defaultGroup during registration may result in downloading large set of policies by the PDP Engine and then filter out which one to use.

    Instead of that if the PDP Engine comes with an identifier which makes PAP identify the PDP Group and then assign it to that group, will result in only the policies assigned to that group being downloaded by the PDP Engine.

    1. The only reason we had a default group was so the policies that are deployed are stored somewhere, as well as when a PDP registers with the PAP there was some group already created for it to move to. But we can re-visit this as I discussed today with Liam Fallon -

      1) Perhaps you cannot deploy a policy with a PDP group (and thus a PDP) being created. eg. return an error

      2) PDP's when they register for the first time, have some PDP group specified in them. The group would then be created to support them.

      I think the above would work without any requirements for a "default" group to always be there. It was actually painful programming that assumption on the XACML side.

  13. I was trying to pen down the component level architecture diagram of the above. Please have a look and let me know your thoughts.

    This will help us in identifying the responsibilities of the individual component. Which in turn will help us in designing the component.

    1. Ok not bad - we are missing the Policy API for Decisions on the Run Time side- those get mapped (at least for now) to the XACML PDP. They are completely different from the "R" on the design time side. "R" on the design time is returning policies as they exist in the DB, not as they are enforced by a decision.

  14. Trying to get knocked down the basic PDP initialization (and registration with the PAP) in a little more level of detail for development to proceed.    This proposal omits some of the lifecycle APIs as to shorten the discussion.    We can iterate over it as we mature it.

    1. PDP gets instantiated.

    2. PDP → PAP Registration   

    First step is for the PDP to register its presence with the PAP at its service endpoint https://policy-pap.   

    The PDP will present to the PAP the following triple:  <identity, pdp-type, capabilities*> where identity is the host name, and capabilities is what this PDP is able to do.   To keep the discussion simple capabilities could be considered a synonym for the concept of domains, although  we may want to keep it separate, with capabilities being an extension of domains.   Regardless, a PDP could also support multiple capabilities or domains.

    The PAP after being presented with the PDP triple <identity, pdp-type, capabilities*> will assign the PDP to a PDP Group and will return the following bootstrapping information:  <pdp-group, policies*>.     The PAP will always return a pdp-group that would be part of the PDP's identity.    The pdp-group identifier is necessary by the PDP for group communication, typically in the direction PAP → PDP, for example to add a policy of policy-type to the group, where there could be multiple recipients.    This also suggests to use DMaaP capabilities as a multicast channel for group based operations.   

    The PAP in its very essence is the resource allocation problem of distributing policies to PDPs, according to some strategy, Therefore, the policies* value in the response pair returned by the PAP would be determined from contextual information and the capabilities* value from the previous paragraph.  The PAP would come up with the initial set of policies for the newly PDP to enforce (policies*).    This is the executable policy set in the architecture terminology.   If the executable policy sets could be considered immutable one possibility is to pass a reference to the set, for example maven coordinates.    The other option is to plainly provide them as part of the response.

    3. PDP operation:

    The PDP is equipped now with its initial sets of policies and group identity.    It will listen in the PAP <-> PDP group channel (let's call it for example PDP-GROUP DMaaP topic) for operations addressed to the group it is a member of.   For example when a policy is to be added to the group, the PAP will place the triple <pdp-group, pdp-type, policy> in this channel.

    1. I agree, this is how I see it working as well.

      I would hope as well that the code to handle all this on the PDP side could be some common for all PDP types. Each PDP will then implement a Java interface that allows them to use the common code.

      1. Yes - i agree with this also. As we briefly mentioned today in the policy meeting, we need to determine how the K8S charts can help align this.