A service unit's direct action is entirely defined by its charm's hooks. Hooks
are executable files in a charm's
hooks directory; hooks with particular names
(see below) will be invoked by the juju unit agent at particular times, and
thereby cause changes to the world.
Whenever a hook-worthy event takes place, the unit agent first checks whether that hook is being debugged, and if so hands over control to the user. Otherwise, it tries to find a hook with precisely the right name. If the hook doesn't exist, the agent continues without complaint; if it does, it is invoked without arguments in a specific environment, and its output is written to the unit agent's log. If it returns a non-zero exit code, the agent enters an error state and awaits user intervention.
The agent will also enter an error state if the unit agent process is aborted during hook execution.
There are two types of hooks, described in more detail in the following sections.
Note: None of the unit or relation hooks are required; if you don't implement a hook, it just doesn't get run. When a hook event occurs, Juju will look for the corresponding hook file to execute, but if it finds none, will continue running without generating an error.
There are 5 "unit hooks" with predefined names that can be implemented by any charm:
For every relation defined by a charm, an additional 4 "relation hooks" can be implemented, named after the charm relation:
install runs just once, before any other hook. It should be used to perform
one-time setup operations only.
config-changed runs in several different situations.
- immediately after "install"
- immediately after "upgrade-charm"
- at least once when the unit agent is restarted (but, if the unit is in an error state, it won't be run until after the error state is cleared).
- after changing charm configuration using a command line interface
It cannot assume that the software has already been started; it should not start stopped software, but should (if appropriate) restart running software to take configuration changes into account.
start runs immediately after the first
config-changed hook. It should be
used to ensure the charm's software is running. Note that the charm's software
should be configured so as to persist through reboots without further
intervention on juju's part.
upgrade-charm runs immediately after any upgrade operation that does not itself interrupt an existing error
state. It should be used to reconcile local state
written by some other version of the charm into whatever form it needs to take
to be manipulated by the current version.
While the forced upgrade functionality is intended as a developer tool, and is
not generally suitable for end users, it's somewhat optimistic to depend on the
functionality never being abused. In light of this, if you need to run an
upgrade-charm hook before your other hooks will work correctly, it may be wise
to preface all your other hooks with a quick call to your (idempotent)
stop runs immediately before the end of the unit's destruction sequence. It
should be used to ensure that the charm's software is not running, and will not
start again on reboot.
This hook is called when a service removal is requested by the client. It should implement the following logic:
- Stop the service
- Remove any files/configuration created during the service lifecycle
- Prepare any backup(s) of the service that are required for restore purposes.
update-status provides constant feedback to the user about the status of the
service the charm is modeling. The charm is run by Juju at regular intervals,
and gives authors an opportunity to run code that gets the "health" of the
service or services.
Units will only participate in relations after they're been started, and before they've been stopped. Within that time window, the unit may participate in several different relations at a time, including multiple relations with the same name.
To illustrate, consider a database service that will be used by multiple client services. Units of a single client service will surely want to connect to, and use, the same database; but if units of another client service were to use that same database, the consequences could be catastrophic for all concerned.
If juju respected the
limit field in relation metadata, it would be possible to work around this, but it's not a high-
priority bug: most provider services
should be able to handle multiple requirers anyway; and most requirers will
only be connected to one provider anyway.
When a unit running a given charm participates in a given relation, it runs at least three hooks for every remote unit it becomes aware of in that relation.
[name]-relation-joined is run once only, when that remote unit is first
observed by the unit. It should be used to
relation-set any local unit
settings that can be determined using no more than the name of the joining unit
and the remote
private-address setting, which is always available when the
relation is created and is by convention not deleted.
You should not depend upon any other relation settings in the -joined hook because they're not guaranteed to be present; if you need more information you should wait for a -changed hook that presents the right information.
[name]-relation-changed is always run once, after
-joined, and will
subsequently be run whenever that remote unit changes its settings for the
relation. It should be the only hook that relies upon remote relation
relation-get, and it should not error if the settings are
incomplete: you can guarantee that when the remote unit changes its settings,
the hook will be run again.
The settings that you can get, and that you should set, are determined by the relation's interface.
[name]-relation-departed is run once only, when the remote unit is known to be
leaving the relation; it will only run once at least one
-changed has been
run, and after
-departed has run, no further
-changed hooks will be run.
This should be used to remove all references to the remote unit, because there's
no guarantee that it's still part of the system; it's perfectly probable
(although not guaranteed) that the system running that unit has already shut
When a unit's own participation in a relation is known to be ending, the unit agent continues to uphold the ordering guarantees above; but within those constraints, it will run the fewest possible hooks to notify the charm of the departure of each individual remote unit.
Once all necessary
-departed hooks have been run for such a relation, the unit
agent will run the final relation hook:
[name]-relation-broken indicates that the current relation is no longer valid,
and that the charm's software must be configured as though the relation had
never existed. It will only be called after every necessary
-departed hook has
been run; if it's being executed, you can be sure that no remote units are
currently known locally.
It is important to note that the
-broken hook might run even if no other units
have ever joined the relation. This is not a bug: even if no remote units have
ever joined, the fact of the unit's participation can be detected in other hooks
relation-ids tool, and the
-broken hook needs to execute to give the
charm an opportunity to clean up any optimistically-generated configuration.
And, again, it's important to internalise the fact that there may be multiple
runtime relations in play with the same name, and that they're independent: one
-broken hook does not mean that every such relation is broken.
Independent of the nuts and bolts, though, good hooks display a number of useful high-level properties:
- They are idempotent: that is to say that there should be no observable difference between running a hook once, and running it N times in a row. If this property does not hold, you are likely to be making your own life unnecessarily difficult: apart from anything else, the average user's most likely first response to a failed hook will be to try to run it again (if they don't just skip it).
They are easy to read and understand. It's tempting to write a single file that does everything, and which just calls different functions internally depending on the value of
argv, and to symlink that one file for every hook; but such structures quickly become unwieldy. The time taken to write a library, separate from the hooks, is very likely to be well spent: it lets you write single hooks that are clear and focused, and insulates the maintainer from irrelevant details.
Where possible, they reuse common code already written to ease or solve common use cases.
- They do not return errors unless there is a good reason to believe that they cannot be resolved without user intervention. Doing so is an admission of defeat: a user who sees your charm returning an error state is unlikely to have the specific expertise necessary to resolve it. If you have to return an error, please be sure to at least write any context you can to the log before you do so.
- They write only very sparingly to the charm directory.