This module serves as the basis for creating charms and relation implementations using the reactive pattern.


The pattern is “reactive” because you use @when and similar decorators to indicate that blocks of code react to certain conditions, such as a relation reaching a specific state, a file changing, certain config values being set, etc. More importantly, you can react to not just individual conditions, but meaningful combinations of conditions that can span multiple hook invocations, in a natural way.

For example, the following would update a config file when both a database and admin password were available, and, if and only if that file was changed, the appropriate service would be restarted:

@when('db.database.available', 'admin-pass')
def render_config(pgsql):
    render_template('app-config.j2', '/etc/app.conf', {
        'db_conn': pgsql.connection_string(),
        'admin_pass': hookenv.config('admin-pass'),

def restart_service():

if __name__ == '__main__':

Structure of a Reactive Charm

The structure of a reactive charm is similar to existing charms, with the addition of reactive directory and the relations directory under hooks:

├── metadata.yaml
├── reactive
│   └──
└── hooks
    ├── pgsql-relation-changed
    └── relations
        └── pgsql
            ├── interface.yaml

The hooks will need to call reactive.main(), and the decorated handler blocks can be placed in any file under the reactive directory. Thus, pretty much all of your hooks will end up contain little more than:

#!/usr/bin/env python
from charms.reactive import main

The relations directory will contain any interface layer implementations that your charm uses.

If you are building a charm with layers, as is recommended, both the hooks and relations directories will be automatically managed for you by your base and interface layers, so you can focus on writing handlers under the reactive directory.

Discovery and Dispatch of Reactive Handlers

Reactive handlers are loaded from any file under the reactive directory, as well as any relation stubs you are using. Handlers can be decorated blocks in Python, or executable files following the ExternalHandler protocol. Handlers can be split amongst several files, which is particularly useful for layers, as each layer can define its own file containing handlers so as not to conflict with files from other layers.

Once all of the handlers are loaded, all @hook handlers will be executed, in a non-determined order. In general, only one layer or relation stub should have a matching @hook block for each hook, which should then set appropriate semantically meaningful states that the other layers can react to. If there are multiple handlers that match for a given hook, there is no guarantee which order they will execute in. Hook handlers should live in the layer that is most appropriate for them. The base or runtime layer will probably handle the install and upgrade hooks, relation stubs will handle all of the relation hooks, etc.

After all of the hook handlers have run, other handlers are dispatched based on the states set by the hook handlers and any states from previous runs. States can be thought of as persistent events. Various hook invocations can each set their appropriate states, and the reactive handlers will be triggered when all of the appropriate states are set, regardless of when and in which order they are each set.

All handlers are tested and matching handlers queued before invoking the first handler. Thus, states set by a handler will not trigger new matching handlers until after all of the current set of matching handlers are done. This allows you to ensure some ordering of otherwise non-determined handler invocation by chaining states (e.g., handler_A sets state_B, which triggers handler_B which then sets state_C, which triggers handler_C, and so on).

Note, however, that removing a state causes the remaining set of matched handlers to be re-tested. This ensures that a handler is never invoked when the state is no longer active.

Relation Stubs

A big part of the reactive pattern is the use of relation stubs. These are classes, based on RelationBase, that are responsible for managing the conversation with remote services or units and informing the charm when the conversation has reached key points, called states, upon which the charm can act and do useful work. They allow a single interface author to create code to handle both sides of the conversation, and to expose a well-defined API to charm authors.

Relation stubs allow charm authors to focus on implementing the behavior and resources that the relation provides, while the interface author focuses on the communication necessary to get that behavior and resources between the related services. In general, the author of the charm that provides a particular interface is likely to be the interface author that creates both the provides and requires sides of the relation. After that, charm authors that wish to make use of that interface can just re-use the existing relation stub.

Non-Python Reactive Handlers

Reactive handlers can be written in any language, provided they conform to the ExternalHandler protocol. In short, they must accept a --test and --invoke argument and do the appropriate thing when called with each.

There are helpers for writing handlers in bash. For example:


@when 'db.database.available' 'admin-pass'
function render_config() {
    db_conn=$(state_relation_call 'db.database.available' connection_string)
    admin_pass=$(config-get 'admin-pass')
    charms.reactive render_template 'app-config.j2' '/etc/app.conf'

@when_not 'db.database.available'
function no_db() {
    status-set waiting 'Waiting on database'

@when_not 'admin-pass'
function no_db() {
    status-set blocked 'Missing admin password'

@when_file_changed '/etc/app.conf'
function restart_service() {
    service myapp restart


Reactive API Documentation


This is the main entry point for the reactive framework. It calls discover() to find and load all reactive handlers (e.g., @when decorated blocks), and then dispatch() to trigger hook and state handlers until the state settles out. Finally, unitdata.kv().flush is called to persist the state.

Parameters:relation_name (str) – Optional name of the relation which is being handled.