Expressed as compactly as possible, the Provisioner is responsible for making sure that non-Dead machine entities in state have agents running on live instances; and for making sure that Dead machines, and stray instances, are removed and cleaned up.

However, the choice of exactly what we deploy involves some subtleties. At the Provisioner level, it's simple: the series and the constraints we pass to the Environ.StartInstance come from the machine entity. But how did they get there?

Individual charms are released for different possible target series; juju should guarantee that charms for series X are only ever run on series X. Every service, unit, and machine has a series that's set at creation time and subsequently immutable. Units take their series from their service, and can only be assigned to machines with matching series.

Subordinate units cannot be assigned directly to machines; they are created by their principals, on the same machine, in response to the creation of subordinate relations. We therefore restrict subordinate relations such that they can only be created between services with matching series.

Constraints are stored for models, services, units, and machines, but unit constraints are not currently exposed because they're not needed outside state, and are likely to just cause trouble and confusion if we expose them.

From the point of a user, there are model constraints and service constraints, and sensible manipulations of them lead to predictable unit deployment decisions. The mechanism is as follows:

• when a unit is added, the current model and service constraints are collapsed into a single value and stored for the unit. (To be clear: at the moment the unit is created, the current service and model constraints will be combined such that every constraint not set on the service is taken from the model (or left unset, if not specified at all).
• when a machine is being added in order to host a given unit, it copies its constraints directly from the unit.
• when a machine is being added without a unit associated -- for example, when adding additional controllers -- it copies its constraints directly from the model.

In this way the following sequence of operations becomes predictable:

    $ juju deploy --constraints mem=2G wordpress
    $ juju set-constraints --application wordpress mem=3G
    $ juju add-unit wordpress -n 2

...in that exactly one machine will be provisioned with the first set of constraints, and exactly two of them will be provisioned using the second set. This is much friendlier to the users than delaying the unit constraint capture and potentially suffering subtle and annoying races.

Subordinate units cannot have constraints, because their deployment is controlled by their principal units. There's only ever one machine to which that subordinate could (and must) be deployed, and to restrict that further by means of constraints will only confuse people.

Placement is the term given to allocating a unit to a specific machine. This is achieved with the --to option in the deploy and add-unit commands.

In addition, it is possible to specify directives to add-machine to allocate machines to specific instances:

• in a new container, possibly on an existing machine (e.g. add-machine lxc:1)
• by using an existing host (i.e. add-machine ssh:user@host)
• using provider-specific features (e.g. add-machine zone=us-east-1a)

At the time of writing, the currently implemented provider-specific placement directives are:

• Availability Zone: both the AWS and OpenStack providers support zone=<zone>, directing the provisioner to start an instance in the specified availability zone.
• MAAS: <hostname> directs the MAAS provider to acquire the node with the specified hostname.

For Juju providers that know about Availability Zones, instances will be automatically spread across the healthy availability zones to maximise service availability. This is achieved by having Juju:

• be able to enumerate each of the availability zones and their current status,
• calculate the "distribution group" for each instance at provisioning time.

The distribution group of a nascent instance is the set of instances for which the availability zone spread will be computed. The new instance will be allocated to the zone with the fewest members of its group.

Distribution groups are intentionally opaque to the providers. There are currently two types of groups: controllers and everything else. controllers are always allocated to the same distribution group; other instances are grouped according to the units assigned at provisioning time. A non-controller instance's group consists of all instances with units of the same services.

At the time of writing, there are currently three providers providers supporting automatic availability zone spread: Microsoft Azure, AWS, and OpenStack. Azure's implementation is significantly different to the others as it contains various restrictions relating to the imposed conflation of high availability and load balancing.

The AWS and OpenStack implementations are both based on the provider/common.ZonedEnviron interface; additional implementations should make use this if possible. There are two components:

• unless a placement directive is specified, the provider's StartInstance must allocate an instance to one of the healthy availability zones. Some providers may restrict availability zones in ways that cannot be detected ahead of time, so it may be necessary to attempt each zone in turn (in order of least-to-most populous);
• the provider must implement state.InstanceDistributor so that units are assigned to machines based on their availability zone allocations.

In the light of time pressure, a unit assigned to a machine that has not been provisioned can be removed directly by calling juju destroy-unit. Any provisioning error can thus be "resolved" in an unsophisticated but moderately effective way:

    $ juju destroy-unit borken/0

...in that at least broken units don't clutter up the service and prevent its removal. However:

    $ juju destroy-machine 1

...does not yet cause an unprovisioned machine to be removed from state (whether directly, or indirectly via the provisioner; the best place to implement this functionality is not clear).

[TODO: figure this out; not yet implemented, somewhat speculative... in particular, use of "resolved" may be inappropriate. Consider adding a "retry" CLI tool...]

When the provisioner fails to start a machine, it should ensure that (1) the machine has no instance id set and (2) the machine has an error status set that communicates the nature of the problem. This must be visible in the output of juju status; and we must supply suitable tools to the user so as to allow her to respond appropriately.

If the user believes a machine's provisioning error to be transient, she can do a simple juju resolved 14 which will set some state to make machine 14 eligible for the provisioner's attention again.

It may otherwise be that the unit ended up snapshotting a service/environ config pair that really isn't satsifiable. In that case, the user can try (say) juju resolved 14 --constraints "mem=2G cpu-power=400", which allows her to completely replace the machine's constraints as well as marking the machine for reprovisioning attention.