Jump to: navigation, search

OpenDaylight Controller:MD-SAL:MD-SAL Document Review:Config SubSystem

WHAT IS IT

The config subsystem is an activation and configuration framework, roughly similar to what Blueprint would do should it also handle statistics and allow for run-time rewiring.

Config subsystem allows the developers to

  • expose configuration and service dependencies of their application,
  • utilize service and configuration injection,
  • expose runtime statistics of their application,
  • leave the lifecycle management of the instances of their application to the config subsystem,
  • package their application into one or more configurable Modules* (one bundle, Multiple configurable modules in it).

It enables the operators and endusers to

  • create new instances of applications at startup or while the controller is running,
  • destroy and reconfigure running applications in a safe and transactional manner,
  • collect statistics of running applications,
  • and to invoke rpcs on Module instances.

* Module is a generic, manageable unit recognized by config subsystem also suitable for JMX.

The config subsystem is a home grown solution and in its core it is just a manager built around JMX registry that enables its users to change Modules in that registry(and their configuration) in a safe manner.

The config subsystem uses yang modeling language as a language for modeling the configuration, dependencies and state data for Modules.

Note: The word module can mean a lot of different things and (unfortunately) its also the element of config subsystem. For this reason we will try to distinguish the Modules from config subsystem by using capital M.

ADDITIONAL COMPONENTS

In addition to the core manager around JMX, there are a few components that fall into(or are coupled with) the config subsystem:

  • Netconf northbound interface - Northbound NETCONF server tied to config subsystem.
  • Yang jmx config generator - Yang-maven-plugin compatible code generator, that generates specific Module interfaces and stub implementations from YANG models.
  • Config persister - Responsible for persisting configuration changes in the controller and pushing initial configuration changes to the controller on startup.
  • Wrapper Modules for commonly used resources - Configurable Modules generated for common resources such as Threadpools to allow configurability, manageability and reusability for these resources.

Detailed information about the components forming config subsystem can be found at the Component map section. In addition, there is a page briefly describing the design of config subsystem (design for netconf subsystem can be found at design of netconf subsystem).

INTERFACES TO CONFIG SUBSYSTEM

Config subsystem is accessible in 4 ways:

  • NETCONF northbound
  • RESTCONF northbound (via a loopback netconf connection instatiated by default using initial config file)
  • JMX (as JMX is the core of config subsystem)
  • From the code

ASSUMED KNOWLEDGE

Users of the config subsystem should consider consulting following resources:

FOR DEVELOPERS

  • YANG - Developers of applications use YANG language in order to describe the configuration, runtime state and dependencies of their applications.
  • JMX (optional) - JMX is used as a core of config subsystem and in order to fully understand the config subsystem (especially config-manager) one should understand how JMX works.
  • Tutorial on how to develop "config subsystem aware" applications - To help application developers start with developing with config subsystem.

FOR OPERATORS

  • RESTCONF/NETCONF/JMX(optional) - These interfaces can be used to read/write data from config subsystem and thus spawn, reconfigure or destroy Modules managed by config subsystem.
  • XML/JSON - RESTCONF enables its users to read/write data using XML or JSON, NETCONF requires usage of XML and JMX requires the use of GUI tool such as jconsole/jvisualvm (parts of standard java JDK).

FOR ENDUSERS

Same knowledge assumed as for operators.

ASSUMED ENVIRONMENT

(if applicable)

WHY DO WE HAVE IT?

  • pain points addressed
  • problems solved

PAINPOINTS

  • Need a general understanding and description of the Config Subsystem. Where to find a good description?
  • What is the “Config Subsystem” and what does it do?
  • Multiple ways of handling config in ODL (via Karaf and via Config Subsystem’s XML files)
  • How does Config Subsystem relate to OSGi bundle dependency resolution?
    • It is not tied to it. Rather then following the whiteboard pattern, where the consumers have to take of filtering, the service instances are directly injected. Services are not imported from the OSGi registry, but all services from config subsystem are published into it.
  • Requires a Yang file to describe the model and an XML file to instantiate the model
  • How do the config xml files (eg, 00-netty.xml) in the configuration/initial directory relate to the Config Subsystem?
    • They are the "initial wiring of the controller". Their format is not tie to the config subsystem, but rather to the config-netconf-persister (e.g. the NETCONF-specific persistence module). They are read and pushed at startup.
  • What is the <required-capabilities> element in the config xml file (eg, 00-netty.xml) used for?
    • These come from NETCONF. They list the models which need to be advertised to be present before the configuration file is attempted to be pushed. Usually features which we reference are listed there -- so that we do not even try to perform a reconfiguration when we know we do not have enough ingredients.
  • What is the preferred way to “inject” references into OSGi components?
  1. Define md-sal dependencies in yang file which auto-generated Module and ModuleFactory classes. These auto-generated classes take care of the injection.
  2. Use an Activator class
  3. Use Declarative Services with Dependency Injection using annotations in the code (eg, @Component, @Reference)
  • There is a lot of documentation that shows you how to do something. BUT…we need documentation that describes “why” we are doing something. When something doesn’t work, you need to understand why something was done to figure out what the problem is.

HOW DOES IT WORK?

This chapter covers basic configuration and use tasks with focus on technologies in use e.g. JMX, OSGi.

CONFIGURATION PROCESS

Configuration process contains details on application development, Module spawning/reconfiguration process and configuration persisting/pushing.

Phase 1:Make application config subsystem aware

First step is to get the application recognized by config subsystem. Assuming there already is an application ready to be included into config subsystem, these are the steps developers need to follow:

  1. Write YANG model describing the interface(in config subsystem called service) and implementation(in config subsystem called Module) of your application.
    • These yang models might be included inside src/main/yang folder of a maven-module for an application or inside a completely separate maven-module that will only deal with configuration for a particular application.
    • YANG model should contain the definition of one or more services and one or more Modules for those services.
    • YANG model needs to depend on config.yang module provided by config-api bundle inside config subsystem, therefore depend on the config-api bundle on maven dependencies level.
    • YANG model config.yang defines the basic structures necessary for the definition of new Modules and services.
  2. Configure the yang-maven-plugin in pom.xml inside maven-module containing YANG models for config subsystem.
  3. Execute maven build process.
  4. Yang-maven-plugin with code generator from config subsystem will generate 2 classes inside src/main/java folder for each module defined in the YANG schemas.
    • FooModuleFactory - Simple factory for Module classes
    • FooModule - JMX compatible wrapper class that wraps an instance of a Module(could be an application or its part). Users are expected to modify this class
    • Additional classes generated in target folder which are not meant for the developer(details in following sections) since these classes are regenerated with every build unlike the FooModuleFactory and FooModule
    • All generated classes are basically the binding between application and config subsystem
  5. Implement the createInstance() method inside the FooModule class to instantiate your application.
    • Getters are used to retrieve the dependencies and configuration needed for application instantiation

Note: Detailed step by step guide can be found in the education section.

Phase 1.1: State data provisioning from Modules

Besides configuration and dependencies, the yang files for config subsystem are used to model read-only state data provided from application (e.g. statistics). Following the steps from previous phase you need to:

  1. Implement interface FooRuntimeMXBean.
    • Generated inside target folder
    • Contains getters for state data modeled in the YANG models from previous phase
  2. In the createInstance() method inside the FooModule register the implementation of FooRuntimeMXBean using injected instance of RootRuntimeBeanRegistrator.

Note: This phase is entirely optional as provision of state data is optional for modules. Note 2: It is also possible to define rpcs along with state data in the yang schemas.

Phase 2:Config subsystem Modules discovery

After the bundles with application and config subsystem bindings are ready, they need to be included in the ODL distribution and picked up by config subsystem at startup. Config subsystem detects "config subsystem aware" bundles by:

  1. Config-manager(in its activator) starts a bundle tracker and listens to "Bundle is active" events.
    • By listening on active, config subsystem is guaranteed that all the dependencies(bundles) of a bundle are already resolved.
    • If all the dependencies are resolved, all YANG model dependencies are resolved as well, since the models are part of those dependencies(bundles).
  2. For every new bundle it looks for generated Module factories.
    • A simple text file: META-INF/services/org.opendaylight.controller.config.spi.ModuleFactory is used to find the generated Module factories
    • The text files contain fully qualified names of all factories inside the bundle
    • The file is added by code generator for config subsystem within yang-maven-plugin while generating config subsystem bindings
  3. For every new bundle it also looks for YANG models to create an instance of a SchemaContext.
    • The SchemaContext is later used to resolve configuration attributes such as identities and also for mapping between NETCONF interface and config subsystem
    • SchemaContext is built using resources from yangtools such as: ModuleInfoBackedContext and ModuleInfoBackedContext
  4. Config-manager creates and stores a singleton instance of each Module factory
    • This instance will be later used to create one or more instances of Modules
    • This instance is also registered in the OSGi service registry
    • Instance of ConfigRegistry is responsible for managing the factories and their instances
    • ConfigRegistry itself is also exposed into OSGi
Phase 3:Application instantiation

After the distribution is fully started and all bundles scanned, it is possible to instantiate available Modules using RESTCONF, NETCONF or JMX.

  1. ConfigRegistry is the core class of config-manager and is responsible for transactional configuration management.
    • Application instantitation is considered a configuration change.
  2. Transaction from ConfigRegistry is started.
  3. New Module instance is created in the transaction (using stored instance of Module factory for a particular Module).
  4. Attributes and dependencies are set for the new instance of the module.
  5. Transaction resolves and validates the dependencies and also other configuration parameters.
    • Users can provide additional validation of configuration attributes inside the generated Module class in method: customValidation()
    • Dependencies are checked for correct dependency type in the module and the dependency graph is searched for loops or missing dependencies (responsibility of DependencyResolver in config-manager)
  6. Transaction is committed.
    • All configuration attributes and dependencies are injected into to Module instance
    • Method createInstance() is called for the Module instance
    • User code inside createInstance() method should instantiate the application and pass all relevant attributes and dependencies to it
    • Returned instance is registered into JMX and optionally into OSGi service registry

Note: Detailed how-to guides are included in education section.
Note 2: Detailed overview of configuration process is also included in education.

Phase 4: Application reconfiguration

After a Module was successfully instantiated, it can be reconfigured at any point, the reconfiguration is very similar to instantiation:

  1. Transaction from ConfigRegistry is started.
  2. The configuration of a Module identified by its type and name is changed.
  3. New module instance is created in the transaction.
  4. Method createInstance() is not called yet.
  5. Initially, method boolean:canReuseInstance(oldModuleInstance) inside generated Module class is called to make the decision whether we can reuse old instance of a Module (typically when no configuration was changed, but this behavior can be overridden by user to modify the default implementation)
    • The logic of reuse and new instance creation is also generated but is located in an abstract class in the target folder. This abstract class is the parent of Module class available to the user and its purpose is to hide general logic common for all Modules. This abstract class however provides some non-final methods, which can be overridden in order to modify the default behavior e.g. reuse of instances.
  6. If old instance can be reused, then the old instance of Module is reused and the new one is abandoned.
    • The transaction can be committed at this point.
  7. If the old instance cannot be reused, then the old Module is abandoned and its wrapped instance(the application it manages) is closed.
    • Method createInstance() is called for the Module instance.
    • The process continues as described in previous phase with newly spawned instances.
    • JMX registration is recreated as the new instance of Module is registered.
Phase: Initial configuration push

In order to automate startup of infrastructural services, plugins and applications, a config-pusher mechanism was implemented. It is part of a config-persister component and detailed documentation can be found at: Initial configuration for ODL (// TODO might need an update for karaf based distribution).

Initial configuration push just consists of a few edit-config rpcs sent to the northbound NETCONF interface for config subsystem. It does not differ from a user sending edit-configs via NETCONF. The format of files with initial configuration was chosen as XML since XML is the format of NETCONF rpcs and allows for trivial construction of edit-config rpcs, since the content is only copied from the files into the rpc. No transformation is needed in such case.

Note: The NETCONF messages with initial config are not transferred over network, just inside the JVM

Required capabilities in files with initial config The whole environment in ODL is fully dynamic as it uses OSGi container internally. This implies the dynamic nature of config subsystem as well as its NETCONF northbound interface. Bundles and features can appear at any time and config subsystem needs to pick them up as they come and go. These bundles may contain YANG models and the binding classes for config subsystem. Before a bundle is installed and picked up, initial config file that tries to spawn a Module defined in that bundle cannot be pushed. Config-pusher has to wait with such config files until required bundles are loaded. The required capabilities listed in the files ensure this behavior as config pusher waits until NETCONF server lists the required YANG models as capabilities.

The required-capabilities in initial config files ensure that config subsystem is ready to spawn required modules in the dynamic environment of OSGi

Config push in karaf

This mechanism was slightly updated for the Karaf distribution. Initial config files are not picked up from file-system but rather from features containing bundles with initial config files. A hook for an install feature event was implemented to scan incoming features. Config-persister is an extensible component that can be configured to use any storage-adapter internally to load initial configuration and the feature-storage-adapter was implemented in this case.

Phase: Configuration persisting

Besides the initial configuration push, config-persister also stores current configuration snapshot from ODL while its running. After every change of configuration, a notification is sent to config-persister with current snapshot to store. Notifications are carried over JMX (JMX can be used easily to transfer notifications). The notifications are generated in the NETCONF interface to config subsystem not config subsystem itself. The reason for this is that inside NETCONF interface there already is a snapshot of current configuration in an easily storeable and restorable format: XML. There is a disadvantage to this approach: Changed configuration would not be stored using config-persister in case users interact with config subsystem directly(from code or JMX).

Stored configuration snapshot will be loaded during the next startup of ODL. This is the mechanism to preserve the changes to the configuration of ODL initiated by operators/users.

TODO: How does the feature-storage-adapter work with persisted configuration ?

YANG IN CONFIG SUBSYSTEM

Primary interfaces for config subsystem are NETCONF and RESTCONF. These interfaces rely on YANG modeling language to describe the data being sent. In addition to data description/definition, YANG models are used to generate the Modules and related config subsystem binding.
YANG modeling language was chosen for config subsystem for 2 reasons:

  1. To fit into the ecosystem of ODL where YANG modeling language is highly used by MD-SAL and applications/plugins.
  2. Considering ODL controller a network device, its management should be exposed in a standardized manner. Management of ODL is handled by the config subsystem and exposed by NETCONF and RESTCONF using YANG as data definition language.
Base Config yang models

Core models for config subsystem(defining top level constructs) are located in config-api component inside config subsystem:

  • config.yang - this model defines following structures:
    • modules - Top level container with a single List node where list entry = Module instance with module specific configuration and state data. Each entry is identified by its module-type and name. Two Choice nodes are used for configuration and state data. Each new module definition then only adds new Case nodes where its configuration and state is defined.
    • services - Top level container with a single List node where list entry = Service reference. Service reference is named reference for an instance of a Module. These references are identified by its service-type and name and used for dependency injection. Module instances cannot be referenced directly as a dependency, instead a service instance has to be used.
    • base identities - Base identities defining module-type and service-type identity. Identities derived from these two specify concrete types of services nad Modules usually defined by users.
    • extensions - Config subsystem specific extensions.
  • rpc-context.yang - constructs and extensions required in case runtime rpc invocation is desired for Module instances.

Note: Detailed information can be found inside the models and in the tutorial/education section.

COMPONENT MAP

HOW DOES IT FIT INTO THE CONTROLLER ARCHITECTURE?

Config subsystem is part of infrastructure of the controller but stands above other infrastructure services(e.g. md-sal) and running applications(e.g. netconf-connector). It gets activated using OSGi Activators and there is only one instance of config subsystem per controller node/JVM. Config-subsystem then manages the modules within the controller (e.g. datastores, md-sal brokers, restconf northbound, netconf southbound connectors etc.).

Everything in the controller that is not config subsystem should be managed by it. An example would be md-sal(datastores, brokers etc.). With config subsystem, it is possible to easily spawn multiple instances of md-sal and wire them with other infrastructure services or applications (currently we are using only one md-sal instance but multiple instances of md-sal could be used in order to isolate/separate concerns within a single controller node and provide different services on top of different md-sal instances e.g. restconf).

Following diagram illustrates the position of config subsystem within the controller (TODO probably not the best diagram, remodel):

Config subsystem in ODL

WHO SHOULD USE IT?

  • Developers, operations, end-user?

WHICH PROJECTS DO/DON'T USE IT?

Config subsystem users:

  • MD-SAL
  • RESTCONF
  • BGPCEP
  • OPENFLOW
  • TOASTER
  • NETCONF CONNECTOR

Config subsystem "resistants"

  • AD-SAL

EDUCATION

This chapter contains links to:

  1. specifications and details for technologies in use,
  2. user guides and tutorials on how to develop with config subsystem and how to use it

LEARNING RESOURCES

Core technologies and config subsystem:

Resources for developers:

Interfaces to config subsystem:

ALTERNATIVES (INVESTIGATED OR POSSIBLE)


Sources:
These resources discuss listed alternatives to config subsystem in detail:

EXAMPLES/TUTORIALS

Meetings/Action Items/Progress