Leadership provides a mechanism whereby multiple units of an application can make use of a single, shared, authoritative source for charm-driven configuration settings.
Every application deployed by Juju is guaranteed to have at most one leader at any time. This is true independent of what the charm author does; whether or not you implement the hooks or use the tools, the unit agents will each seek to acquire leadership, and maintain it while they have it or wait for the current leader to drop out.
If you wish to be notified when your unit's leadership status changes, you should implement the following hooks:
leader-elected(which will run at least once, when the unit is known to be leader and guaranteed to remain so for at least 30s).
leader-settings-changed(which will run at least once when the unit is not guaranteed continued leadership for the next 30s; and also whenever some other unit writes leader settings).
No particular guarantees can be made regarding the timeliness of the
leader-settings-changed hook; it's always possible for the Juju agent itself
to be taken out of commission at the wrong moment and not restarted for a long
Every unit can discover whether it's leader, independent of the hook that's running. There is deliberately no mechanism for discovering which other unit is the leader; such data always risks staleness and opens the door to a lot of unprepossessing race scenarios.
"False"to stdout, and return 0, if the unit is currently leader and can be guaranteed to remain so for 30s. Output can be expressed as
--format yamlif desired. If it returns a non-zero exit code, no inferences regarding true leadership status can be made, but you should generally fail safe and refrain from acting as leader when you cannot be sure.
is-leadertruth is independent of hook sequence. If a unit has been designated leader while mid-hook, it will start to return true; and if a unit were to (say) lose its state-server connection mid-hook, and be unable to verify continued leadership past lease expiry time, it would start to return false.
Every application deployed by Juju also has access to a pseudo-relation over which leader settings can be communicated with the following tools:
leader-setacts much like
relation-set, in that it lets you write string key/value pairs (in which an empty value removes the key), but with the following differences:
- there's only one leader-settings bucket per application (not one per unit)
- only the leader can write to the bucket
- only minions are informed of changes to the bucket
- changes are propagated instantly, bypassing the sandbox
The instant propagation is surprising, and exists to satisfy the use case where shared data can be chosen by the leader at the very beginning of (say) the install hook; by propagating it instantly, any running minions can make use of the data and progress immediately, without having to wait for the leader to finish its hook.
It also means that you can guarantee that a successful
leader-setcall has been reflected in the database, and that all minions will converge towards seeing that value, even if an unexpected error takes down the current hook.
For both these reasons we strongly recommend that leader settings are always written as a self-consistent group (
leader-set foo=bar baz=qux ping=pong, rather than
leader-set foo=bar; leader-set baz=quxetc, in which minions may end up seeing a sandbox in which only
foois set to the "correct" value).
leader-getacts much like relation-get, in that it lets you read string values by key (and expose them in helpful formats), but with the following difference:
- it reads only from the single leader-settings bucket
...and the following key similarity:
- it presents a sandboxed view of leader-settings data.
This is necessary, as it is for relation data, because a hook context needs to present consistent data; but it means that there's a small extra burden on users of
Assuming your own implementations of
can use the two pseudopython snippets below:
def set_shared_settings(): if is_leader(): if !valid_settings(leader_get()): settings = create_settings() leader_set(settings) def get_shared_settings(): settings = leader_get() if !valid_settings(settings): raise WaitingForLeader() return settings
...which can be used as follows:
set_shared_settingsmust be called in
leader-elected, and may be called anywhere else you like (for example, at the very beginning of
install, to cause those settings to be ppropagated to minions as soon as possible).
get_shared_settingsmust be called (and handled!) in
leader-settings-changed, and may also be called at any other time it's convenient; you should always be prepared to catch WaitingForLeader and handle it appropriately (most likely by setting a "waiting" status and exiting without error, to wait for the
leader-settings-changedwhich should arrive soon).
You should be able to use the exact same constructs as above, in the same way;
you just might want to call
set_shared_settings in a few more places. If you
need additional synchronisation, you can use a peer relation to communicate
minions' acknowledgements back to the leader.
Note: peer relation membership is not guaranteed to match current reality at any given time. To be resilient in the face of your application scaling at the same time as you (say) rebalance your application, your leader code will need to use the output of
status-get --application to determine up-to-date membership, and wait for the set of acked units in the peer relation to match that list.
hacluster charm used in our OpenStack deploys will set up corosync and
pacemaker, and may well be relevant to your needs; if that's not a good fit,
Juju's leadership concept is, by choice, relatively fine-grained, to ensure
timely handover of agent-level responsibilities. That's why
guarantees only 30s of leadership; but it's no fun running a separate watchdog
juju-run is-leader every 30s and kill your process when that stops
working (apart from anything else, your
juju-run could be blocked by other
operations, so you can't guarantee a run every 30s anyway).
And we don't plan to allow coarser-grained leadership requests. This is because if one unit could declare itself leader for a day (or even an hour) a failed leader will leave other parts of Juju blocked for that length of time, and we're not willing to take on that cost; the 30-60s handover delay is bad enough already.
So, you'd basically have to implement your own protocol on top of the available
primitives. As a charm author, this is unlikely to be the best use of your time
-- extending the
hacluster charm to cover your use case is likely to be more
Note: this approach is not reliable. It may be good enough for some workloads, but don't use it unless you understand the forces in play and the worst possible consequences for your users...
If you start your long-lived process in
leader-elected, and stop it in
leader-settings-changed, this will usually do what you want, but is
vulnerable to a number of failure modes -- both because hook execution may be
blocked until after the leadership lease has actually expired, and because
total Juju failure could also cause the hook not to run (but leave the workload
In the future, we may implement a
leader-deposed hook, that can run with
stronger timeliness guarantees; but even if we do, it's a fundamentally
unreliable approach. Seriously, if you possibly can, go with the charmed-up