Skip to main content Link Menu Expand (external link) Document Search Copy Copied

Edge development concepts

Open Horizon (OH) is built on the Open Horizon open source software.

With Open Horizon, you can develop any service containers that you want for your edge machines. You can then cryptographically sign the container configuration and publish it. Finally, you can deploy your service containers using a deployment policy or pattern to govern software installation, monitoring, and updating. After you complete these tasks, you can view the Horizon agents and Horizon Agbots forming agreements to collaborate on managing the software lifecycle. These components autonomously manage the software lifecycle details on your Open Horizon edge nodes. Open Horizon can also use policies to autonomously deploy machine learning models. For information on machine learning model deployment, see Model Management System.

The software development process within Open Horizon is focused on maintaining system security and integrity, while greatly simplifying the effort that is required for active software management on your edge nodes. You can build Open Horizon publishing procedures into your continuous integration and deployment pipeline. When the distributed autonomous agents discover published changes in the software or a policy, such as within the Open Horizon deployment pattern or deployment policy, the autonomous agents independently act to update the software and enforce your policies across your entire fleet of edge machines, wherever they are located.


Open Horizon services are the building blocks of edge solutions. Each service contains one or more Docker containers. Each Docker container can in turn contain one or more long-running processes. These processes can be written in almost any programming language, and use any libraries or utilities. However, the processes must be developed for, and run in, the context of a Docker container. This flexibility means that there are almost no constraints on the code that Open Horizon can manage for you. When a container runs, the container is constrained in a secure sandbox. This sandbox restricts access to hardware devices, some operating system services, the host file system, the host edge machine networks and most importantly other services running on the edge node. For information on sandbox constraints, see Sandbox.

The cpu2evtstreams example code consists of a Docker container that uses two other edge services. These edge services connect over local private Docker virtual networks by using HTTP REST APIs. These services are named cpu and gps. The agent deploys each service on a separate private network along with each service that declared a dependency on the service. One network is created for cpu2evtstreams and cpu, and another network is created for cpu2evtstreams and gps. If there is a fourth service in this deployment that is also sharing the cpu service, then another private network is created for just the cpu and the fourth service. In Open Horizon, this network strategy restricts access for services to only the other services that are listed in requiredServices when the other services were published. The following diagram shows the cpu2evtstreams deployment when the pattern runs on an edge node:

Services in a pattern

Note: Set up IBM Event Streams only needed for some examples.

The two virtual networks enable the cpu2evtstreams service container to access the REST APIs that are provided by the cpu and gps service containers. These two containers manage access to operating system services and the hardware devices. Although REST APIs are used, there are many other possible forms of communication that you can use to enable your services to share data and control.

Often the most effective coding pattern for edge nodes involves deploying multiple small, independently configurable, and deployable services. For example, Internet of Things patterns often have low-level services that need access to the edge node hardware, such as sensors or actuators. These services provide shared access to this hardware for other services to use.

This pattern is useful when the hardware requires exclusive access to provide a useful function. The low-level service can properly manage this access. The role of the cpu and gps service containers is similar in principle to that of the device driver software in the host’s operating system, but at a higher level. Segmenting the code into independent small services, some specializing in low-level hardware access, enables a clear separation of concerns. Each component is free to evolve and be updated in the field independently. Third-party applications can also be securely deployed together along with your proprietary traditional embedded software stack by selectively allowing them access to particular hardware or other services.

For example, an industrial controller deployment might be composed of a low-level service for monitoring power usage sensors and other low-level services. These other low-level services can be used to enable control of the actuators for powering the devices that are monitored. The deployment might also have another top-level service container that consumes the services of the sensor and actuator. This top-level service can use the services to alert operators or to automatically power down devices when anomalous power consumption readings are detected. This deployment might also include a history service that records and archives sensor and actuator data, and possibly complete analysis on the data. Other useful components of such a deployment might be a GPS location service.

Each individual service container can be independently versioned and updated with this design. Each individual service might also be reconfigured and composed into other useful deployments without any code changes. If needed, a third-party analytics service can be added to the deployment. This third-party service can be given access to only a particular set of read-only APIs, which prevents the service from interacting with the actuators on the platform.

Alternatively, all of the tasks in this industrial controller example can be run within a single service container. This alternative is not usually the best approach since a collection of smaller independent and interconnected services usually makes software updates faster and more flexible. Collections of smaller services can also be more robust in the field. For more information about how to design a deployment, see Edge-native development practices.


The sandbox in which deployments run restricts access to APIs that are provided by other service containers. Only services that explicitly state dependencies on your services are permitted access. Other processes on the host are unable to access to these services. Similarly, other remote hosts are unable to access any of your services unless your service explicitly publishes a port to the host’s external network interface. The sandbox’s access control restrictions are determined by network addressability, not by an administered access control list. This is accompished by creating virtual networks for each service, and only service containers that are allowed to communicate are connected to the same network. This alleviates the need to configure access control on each edge node.

Services that use other services

Edge services often use various API interfaces that are provided by other edge services to acquire data from them, or to deliver control commands to them. These API interfaces are commonly HTTP REST APIs, like the ones provided by the low-level cpu and gps services in the cpu2evtstreams example. However, those interfaces can really be anything that you want, such as shared memory, TCP, or UDP, and can be with or without encryption. Since these communications typically take place within a single edge node, with messages never leaving this host, often encryption is unnecessary.

As an alternative to REST APIs, you can use a publishing and subscribing interface, such as the interface that is provided by MQTT. When a service provides data intermittently only, a publishing and subscribing interface is usually simpler than repeatedly polling a REST API as the REST APIs can timeout. For example, consider a service that monitors a hardware button, and provides an API for other services to detect whether a button press occurred. If a REST API is used, the caller cannot call the REST API and wait for a reply that would come when the button was pressed. If the button remained unpressed for too long, the REST API would timeout. Instead, the API provider would need to respond promptly to avoid an error. The caller must repeatedly and frequently call the API to be sure not to miss a brief button press. A better solution is for the caller to subscribe to an appropriate topic on a publishing and subscribing service and block. Then, the caller can wait for something to be published, which might occur far in the future. The API provider can take care of monitoring the button hardware and then publish only the state changes to that topic, such as button pressed, or button released.

MQTT is one of the more popular publishing and subscribing tools that you can use. You can deploy an MQTT broker as an edge service, and have your publisher and subscriber services require it. MQTT is also frequently used as a cloud service. The IBM Watson IoT Platform, for example, uses MQTT to communicate with IoT devices. For more information, see IBM Watson IoT Platform . Some of the Open Horizon project examples use MQTT. For more information, see Open Horizon examples.

Another popular publishing and subscribing tool is Apache Kafka, which is also frequently used as a cloud service. Apache Kafka, which is used by the cpu2evtstreams example to send data to the IBM Cloud, is also based on Kafka. For more information, see Apache Kafka .

Any edge service container can provide or consume other local edge services on the same host, and edge services provided on nearby hosts on the local LAN. Containers might communicate with centralized systems in a remote corporate or cloud provider data center. As a service author, you determine with whom and how your services communicate. When communicating with cloud provider services, use secrets to contain the authentication credentials as described in Developing Secrets.

You might find it useful to review the cpu2evtstreams example again to see how the example code uses the other two local services. For instance, how the example code specifies dependencies on the two local services, declares and uses configuration variables, and communicates with Kafka. For more information, see cpu2evtstreams example.

Privileged mode services

On a host machine, some tasks can only be performed by an account with root access. The equivalent for containers is privileged mode. While containers generally do not need privileged mode on the host, there are some use cases where it is required. In Open Horizon you have the ability to specify that a service should be deployed with privileged process execution enabled. By default, it is disabled. You must explicitly enable it in the deployment configuration of the respective Service Definition file for each service that needs to run in this mode. And further, any node on which you want to deploy that service must also explicitly allow privileged mode containers. This ensures that node owners have some control over which services are executing on their edge nodes. For an example of how to enable privileged mode policy on an edge node , see privileged node policy . If the service definition or one of its dependencies requires privileged mode, the node policy must also allow privileged mode, or else none of the services will not be deployed to the node. For an in-depth discussion of privileged mode see What is privileged mode and do I need it? .

Service definition

Note: See Conventions used in this document for more information about command syntax.

In every Open Horizon project, you have a horizon/service.definition.json file. This file defines your edge service for two purposes. One of these purposes is to enable you to simulate the running of your service using the hzn dev tool. This tool simulates a real agent environment including the network Sandbox. This simulation is useful for working out any special deployment instructions that you might need, such as port bindings, and hardware device access. The simulation is also useful for verifying communications between service containers on the docker virtual private networks that the agent creates for you. The other reason for this file is to enable you to publish your service to the Horizon exchange. In the provided examples, the horizon/service.definition.json file is either provided for you within the example GitHub repository or is generated by the hzn dev service new command.

Open the horizon/service.definition.json file that contains the Horizon metadata for one of the example service implementations, for example, the cpu2evtstreams .

Every service that is published in Horizon needs to have a name that uniquely identifies it within your organization. The name is placed in the url field and forms a globally unique identifier, when combined with your organization name, and a specific implementation version and hardware arch. For a full description of the service definition see Service Definition. The cpu2evtstreams example exploits some additional features of a basic service definition such as required services and service variables.

The requiredServices section of the horizon/service.definition.json file itemizes any service dependencies that this service uses. The hzn dev dependency fetch tool lets you add dependencies to this list, so you do not need to manually edit the list. After dependencies are added, when the agent runs the container, those other requiredServices are automatically run (for example, when you use hzn dev service start or when you register a node to which this service is deployed). For more information about required services, see Service Definition and cpu2evtstreams.

In the userInput section, you declare the service variables that your Service can consume to configure itself for a particular deployment. You declare variable names, data types, and default values here and you might also provide a human-readable description for each. When you use hzn dev service start or when you register an edge node on which this service is deployed, you need to configure these service variables. The cpu2evtstreams example does this by providing a userinput.json file during node registration. It is also possible to set service variables remotely via the CLI command hzn exchange node update -f <userinput-settings-file>. For more information about service variables see Service Definition and cpu2evtstreams.

The horizon/service.definition.json file also contains a deployment section, toward the end of the file. The fields in this section name each Docker container that implements your logical Service. The name of each container in the services array is the DNS name that other containers use to identify the container on the shared virtual private network. If this container provides a REST API for other containers to consume, you can access this REST API within the consuming container by using curl http://<name>/<your-rest-api-uri>. The image field for each name provides a reference to the corresponding Docker container image, such as within DockerHub or some private container registry. Other fields in the deployment section can be used to configure the container with runtime options that Docker uses to run the container. For more information, see Horizon deployment configuration.

What to do next

For more information about developing edge node code, review the following documentation:

  • Edge-native development practices

    Review the important principles and best practices for developing edge services for Open Horizon software development.

  • Using DockerHub

    With Open Horizon, you can optionally put service containers into the IBM private secure container registry instead of the public Docker Hub. For instance, if you have a software image that includes assets that are not appropriate to include in a public registry, you can use a private Docker container registry like the DockerHub.

  • APIs

    Open Horizon provides RESTful APIs to collaborate and enables your organization’s developers and users to control the components.

  • Updating an edge service with rollback

    Review additional details on how to roll out a new version of an existing edge service and the software development best practices for updating rollback settings in pattern or deployment policies.