Now You're Charming with Layers

Charming with layers is a relatively new approach to writing charms that we, as the Juju community, are begining to use. The idea is to allow better re-use of common code in charms, more well-defined and consistent implementations of interface protocols, and generally allow charm authors to focus on the parts of the charm that is most relevant to their charm and leverage the community for everything else.

Charming with layers is starting to gain momentum. We’re seeing new interface and base layers being added to interfaces.juju.solutions (which can be done by anyone with a Launchpad ID), we’ve had discussions on creating a standardized Java interface layer for providing JREs or JDKs, and I’ve been working this week on creating interface layers for the Apache Hadoop big data charms to make it easier to connect to those charms and to enable us to restructure the core Hadoop charms using layers to make them easier to understand and maintain. So I thought that it would be a good time to write up a short explanation about what the different types of charm layers are, what I think belongs in them, and how they can be used to make writing charms a much more pleasant experience.

Types of Layers

There are three major categories of charm layers:

  • Base, or runtime layers
  • Interface layers
  • Charm layers

Each of these has a distinct role, and it’s important to understand how a charm should be broken up into these types of layers. Generally, a charm will contain one base layer, one charm layer, and one or more interface layers, but it is possible that a charm might include more than one base layer, as well.

Base, or Runtime, Layers

Base layers are layers that other charms can be built on. They contain things that are common to many different charms, and allow charms to reuse that commonality without having to reimplement it each time. Base layers typically are not sufficient on their own to be considered a charm; they likely can’t be built into a deployable charm, and if they can, they’re unlikely to do anything useful. However, other than how they’re used, there’s nothing inherent that distinguishes base layers from charm layers. It is possible for a fully working charm to be written to also be used as a base for other charms, if that use-case makes sense for the charm. It should also be noted that it is common for a base layer to build on another base layer, especially layer-basic.

The most basic example is just that, layer-basic. It provides nothing more than the minimum needed to effectively use layered charms: charms.reactive, charmhelpers, and the skeleton hook implementations that call into the reactive framework. However, the most useful base layers are actually a type of runtime layer. For example, layer-apache-php provides Apache2 and mod-php, as well as mechanisms for fetching and installing a PHP project within that runtime.

Base layers can be written in any language, but must at a minimum provide the reactive framework that glues layers together, which is written in Python. This can be done trivially by building the base layer off of layer-basic.

Interface Layers

Interface layers are perhaps the most confusing type of layer, and are responsible for the communication that transpires over a relation between two services. This type of layer encapsulates a single “interface protocol” and is generally written and maintained by the author of the primary charm that provides that interface. However, it does cover both sides (provides and requires) of the relation and turns the two-way key-value store that are Juju relations under-the-hood into a full-fledged API for interacting with charms supporting that interface.

It is important to note that interface layers do not actually implement either side of the relation. Instead, they are solely responsible for the communication that goes on over the relation, relying on charms on either end to decide what to do with the results of that communication.

A concrete example would be the interface-hadoop-plugin layer for the big data charm ecosystem. It manages the communication between the apache-hadoop-plugin charm, and any charm that would want to use the plugin to connect to the core Hadoop cluster. Note that it does not implement the plugin functionality, it just manages the the API that is exposed by the plugin to the charm using it.

As another example, there is work in progress to define a JRE or JDK interface that would standardize how a charm could provide a JRE to another service and make it easy to choose which vendor’s version of Java that service uses. (This would make it easy to use a version of Java tailored to the specific hardware you happen to be deploying on, for instance.) The various JRE or JDK charms, such as Zulu8 or IBMJavaSDK, would then provide the java relation interface, and various charms, such as Tomcat or WebSphere-Liberty, would require it.

The interface-java layer would not actually install or configure Java, nor would it manage Tomcat or WebSphere. What it would do is provide a consistent API (through reactive states) that would inform the Tomcat or WebSphere charms when a JRE has been attached and provide methods that can be called to send configuration options back to the charm providing Java, as well as an API for the Java charms to receive those configuration options and methods for them to send back relevant information.

Interface layers currently must be written in Python and extend the ReactiveBase class, though they can then be used by any language using the built-in CLI API.

Charm Layers

Building on base and interface layers, charm layers are what actually get turned into charms. This is where the core logic of the charm should go, the logic specific to that individual charm. This layer brings together all the pieces needed to create the charm. It is where most of the charm’s config options will be defined, and where the reactive handlers that do the specific work of the charm will go. It will need to contain the charm’s README, copyright, icon, and so on.

Charm layers should be the most common type of layer, and is what most charm authors will be dealing with. However, the goal is to keep them hyper-focused on just that specific charm’s logic and needs, and to push any commonality into an appropriate base layer. Charm layers should contain as little boilerplate as possible.

Charm layers can be written in any language, and there are helpers for writing them in Bash.