Terraform Resource Instance Change Lifecycle

This document describes the relationships between the different operations called on a Terraform Provider to handle a change to a resource instance.

The resource instance operations all both consume and produce objects that conform to the schema of the selected resource type.

The overall goal of this process is to take a Configuration and a Previous Run State, merge them together using resource-type-specific planning logic to produce a Planned State, and then change the remote system to match that planned state before finally producing the New State that will be saved in order to become the Previous Run State for the next operation.

The various object values used in different parts of this process are:

  • Configuration: Represents the values the user wrote in the configuration, after any automatic type conversions to match the resource type schema.

    Any attributes not defined by the user appear as null in the configuration object. If an argument value is derived from an unknown result of another resource instance, its value in the configuration object could also be unknown.

  • Prior State: The provider's representation of the current state of the remote object at the time of the most recent read.

  • Proposed New State: Terraform Core uses some built-in logic to perform an initial basic merger of the Configuration and the Prior State which a provider may use as a starting point for its planning operation.

    The built-in logic primarily deals with the expected behavior for attributes marked in the schema as “computed”. If an attribute is only “computed”, Terraform expects the value to only be chosen by the provider and it will preserve any Prior State. If an attribute is marked as “computed” and “optional”, this means that the user may either set it or may leave it unset to allow the provider to choose a value.

    Terraform Core therefore constructs the proposed new state by taking the attribute value from Configuration if it is non-null, and then using the Prior State as a fallback otherwise, thereby helping a provider to preserve its previously-chosen value for the attribute where appropriate.

  • Initial Planned State and Final Planned State are both descriptions of what the associated remote object ought to look like after completing the planned action.

    There will often be parts of the object that the provider isn't yet able to predict, either because they will be decided by the remote system during the apply step or because they are derived from configuration values from other resource instances that are themselves not yet known. The provider must mark these by including unknown values in the state objects.

    The distinction between the Initial and Final planned states is that the initial one is created during Terraform Core's planning phase based on a possibly-incomplete configuration, whereas the final one is created during the apply step once all of the dependencies have already been updated and so the configuration should then be wholly known.

  • New State is a representation of the result of whatever modifications were made to the remote system by the provider during the apply step.

    The new state must always be wholly known, because it represents the actual state of the system, rather than a hypothetical future state.

  • Previous Run State is the same object as the New State from the previous run of Terraform. This is exactly what the provider most recently returned, and so it will not take into account any changes that may have been made outside of Terraform in the meantime, and it may conform to an earlier version of the resource type schema and therefore be incompatible with the current schema.

  • Upgraded State is derived from Previous Run State by using some provider-specified logic to upgrade the existing data to the latest schema. However, it still represents the remote system as it was at the end of the last run, and so still doesn't take into account any changes that may have been made outside of Terraform.

  • The Import ID and Import Stub State are both details of the special process of importing pre-existing objects into a Terraform state, and so we'll wait to discuss those in a later section on importing.

Provider Protocol API Functions

The following sections describe the three provider API functions that are called to plan and apply a change, including the expectations Terraform Core enforces for each.

For historical reasons, the original Terraform SDK is exempt from error messages produced when certain assumptions are violated, but violating them will often cause downstream errors nonetheless, because Terraform's workflow depends on these contracts being met.

The following section uses the word “attribute” to refer to the named attributes described in the resource type schema. A schema may also include nested blocks, which contain their own set of attributes; the constraints apply recursively to these nested attributes too.

The following are the function names used in provider protocol version 6. Protocol version 5 has the same set of operations but uses some marginally-different names for them, because we used protocol version 6 as an opportunity to tidy up some names that had been awkward before.

ValidateResourceConfig

ValidateResourceConfig takes the Configuration object alone, and may return error or warning diagnostics in response to its attribute values.

ValidateResourceConfig is the provider's opportunity to apply custom validation rules to the schema, allowing for constraints that could not be expressed via schema alone.

In principle a provider can make any rule it wants here, although in practice providers should typically avoid reporting errors for values that are unknown. Terraform Core will call this function multiple times at different phases of evaluation, and guarantees to eventually call with a wholly-known configuration so that the provider will have an opportunity to belatedly catch problems related to values that are initially unknown during planning.

If a provider intends to choose a default value for a particular optional+computed attribute when left as null in the configuration, the provider must tolerate that attribute being unknown in the configuration in order to get an opportunity to choose the default value during the later plan or apply phase.

The validation step does not produce a new object itself and so it cannot modify the user's supplied configuration.

PlanResourceChange

The purpose of PlanResourceChange is to predict the approximate effect of a subsequent apply operation, allowing Terraform to render the plan for the user and to propagate the predictable subset of results downstream through expressions in the configuration.

This operation can base its decision on any combination of Configuration, Prior State, and Proposed New State, as long as its result fits the following constraints:

  • Any attribute that was non-null in the configuration must either preserve the exact configuration value or return the corresponding attribute value from the prior state. (Do the latter if you determine that the change is not functionally significant, such as if the value is a JSON string that has changed only in the positioning of whitespace.)

  • Any attribute that is marked as computed in the schema and is null in the configuration may be set by the provider to any arbitrary value of the expected type.

  • If a computed attribute has any known value in the planned new state, the provider will be required to ensure that it is unchanged in the new state returned by ApplyResourceChange, or return an error explaining why it changed. Set an attribute to an unknown value to indicate that its final result will be determined during ApplyResourceChange.

PlanResourceChange is actually called twice per run for each resource type.

The first call is during the planning phase, before Terraform prints out a diff to the user for confirmation. Because no changes at all have been applied at that point, the given Configuration may contain unknown values as placeholders for the results of expressions that derive from unknown values of other resource instances. The result of this initial call is the Initial Planned State.

If the user accepts the plan, Terraform will call PlanResourceChange a second time during the apply step, and that call is guaranteed to have a wholly-known Configuration with any values from upstream dependencies taken into account already. The result of this second call is the Final Planned State.

Terraform Core compares the final with the initial planned state, enforcing the following additional constraints along with those listed above:

  • Any attribute that had a known value in the Initial Planned State must have an identical value in the Final Planned State.

  • Any attribute that had an unknown value in the Initial Planned State may either remain unknown in the second or take on any known value that conforms to the unknown value's type constraint.

The Final Planned State is what passes to ApplyResourceChange, as described in the following section.

ApplyResourceChange

The ApplyResourceChange function is responsible for making calls into the remote system to make remote objects match the Final Planned State. During that operation, the provider should decide on final values for any attributes that were left unknown in the Final Planned State, and thus produce the New State object.

ApplyResourceChange also receives the Prior State so that it can use it to potentially implement more “surgical” changes to particular parts of the remote objects by detecting portions that are unchanged, in cases where the remote API supports partial-update operations.

The New State object returned from the provider must meet the following constraints:

  • Any attribute that had a known value in the Final Planned State must have an identical value in the new state. In particular, if the remote API returned a different serialization of the same value then the provider must preserve the form the user wrote in the configuration, and must not return the normalized form produced by the provider.

  • Any attribute that had an unknown value in the Final Planned State must take on a known value whose type conforms to the type constraint of the unknown value. No unknown values are permitted in the New State.

After calling ApplyResourceChange for each resource instance in the plan, and dealing with any other bookkeeping to return the results to the user, a single Terraform run is complete. Terraform Core saves the New State in a state snapshot for the entire configuration, so it'll be preserved for use on the next run.

When the user subsequently runs Terraform again, the New State becomes the Previous Run State verbatim, and passes into UpgradeResourceState.

UpgradeResourceState

Because the state values for a particular resource instance persist in a saved state snapshot from one run to the next, Terraform Core must deal with the possibility that the user has upgraded to a newer version of the provider since the last run, and that the new provider version has an incompatible schema for the relevant resource type.

Terraform Core therefore begins by calling UpgradeResourceState and passing the Previous Run State in a raw form, which in current protocol versions is the raw JSON data structure as was stored in the state snapshot. Terraform Core doesn‘t have access to the previous schema versions for a provider’s resource types, so the provider itself must handle the data decoding in this upgrade function.

The provider can then use whatever logic is appropriate to update the shape of the data to conform to the current schema for the resource type. Although Terraform Core has no way to enforce it, a provider should only change the shape of the data structure and should not change the meaning of the data. In particular, it should not try to update the state data to capture any changes made to the corresponding remote object outside of Terraform.

This function then returns the Upgraded State, which captures the same information as the Previous Run State but does so in a way that conforms to the current version of the resource type schema, which therefore allows Terraform Core to interact with the data fully for subsequent steps.

ReadResource

Although Terraform typically expects to have exclusive control over any remote object that is bound to a resource instance, in practice users may make changes to those objects outside of Terraform, causing Terraform's records of the object to become stale.

The ReadResource function asks the provider to make a best effort to detect any such external changes and describe them so that Terraform Core can use an up-to-date Prior State as the input to the next PlanResourceChange call.

This is always a best effort operation because there are various reasons why a provider might not be able to detect certain changes. For example:

  • Some remote objects have write-only attributes, which means that there is no way to determine what value is currently stored in the remote system.
  • There may be new features of the underlying API which the current provider version doesn't know how to ask about.

Terraform Core expects a provider to carefully distinguish between the following two situations for each attribute:

  • Normalization: the remote API has returned some data in a different form than was recorded in the Previous Run State, but the meaning is unchanged.

    In this case, the provider should return the exact value from the Previous Run State, thereby preserving the value as it was written by the user in the configuration and thus avoiding unwanted cascading changes to elsewhere in the configuration.

  • Drift: the remote API returned data that is materially different from what was recorded in the Previous Run State, meaning that the remote system's behavior no longer matches what the configuration previously requested.

    In this case, the provider should return the value from the remote system, thereby discarding the value from the Previous Run State. When a provider does this, Terraform may report it to the user as a change made outside of Terraform, if Terraform Core determined that the detected change was a possible cause of another planned action for a downstream resource instance.

This operation returns the Prior State to use for the next call to PlanResourceChange, thus completing the circle and beginning this process over again.

Handling of Nested Blocks in Configuration

Nested blocks are a configuration-only construct and so the number of blocks cannot be changed on the fly during planning or during apply: each block represented in the configuration must have a corresponding nested object in the planned new state and new state, or Terraform Core will raise an error.

If a provider wishes to report about new instances of the sub-object type represented by nested blocks that are created implicitly during the apply operation -- for example, if a compute instance gets a default network interface created when none are explicitly specified -- this must be done via separate “computed” attributes alongside the nested blocks. This could be list or map of objects that includes a mixture of the objects described by the nested blocks in the configuration and any additional objects created implicitly by the remote system.

Provider protocol version 6 introduced the new idea of structural-typed attributes, which are a hybrid of attribute-style syntax but nested-block-style interpretation. For providers that use structural-typed attributes, they must follow the same rules as for a nested block type of the same nesting mode.

Import Behavior

The main resource instance change lifecycle is concerned with objects whose entire lifecycle is driven through Terraform, including the initial creation of the object.

As an aid to those who are adopting Terraform as a replacement for existing processes or software, Terraform also supports adopting pre-existing objects to bring them under Terraform's management without needing to recreate them first.

When using this facility, the user provides the address of the resource instance they wish to bind the existing object to, and a string representation of the identifier of the existing object to be imported in a syntax defined by the provider on a per-resource-type basis, which we'll call the Import ID.

The import process trades the user's Import ID for a special Import Stub State, which behaves as a placeholder for the Previous Run State pretending as if a previous Terraform run is what had created the object.

ImportResourceState

The ImportResourceState operation takes the user's given Import ID and uses it to verify that the given object exists and, if so, to retrieve enough data about it to produce the Import Stub State.

Terraform Core will always pass the returned Import Stub State to the normal ReadResource operation after ImportResourceState returns it, so in practice the provider may populate only the minimal subset of attributes that ReadResource will need to do its work, letting the normal function deal with populating the rest of the data to match what is currently set in the remote system.

For the same reasons that ReadResource is only a best effort at detecting changes outside of Terraform, a provider may not be able to fully support importing for all resource types. In that case, the provider developer must choose between the following options:

  • Perform only a partial import: the provider may choose to leave certain attributes set to null in the Prior State after both ImportResourceState and the subsequent ReadResource have completed.

    In this case, the user can provide the missing value in the configuration and thus cause the next PlanResourceChange to plan to update that value to match the configuration. The provider's PlanResourceChange function must be ready to deal with the attribute being null in the Prior State and handle that appropriately.

  • Return an error explaining why importing isn't possible.

    This is a last resort because of course it will then leave the user unable to bring the existing object under Terraform‘s management. However, if a particular object’s design doesn't suit importing then it can be a better user experience to be clear and honest that the user must replace the object as part of adopting Terraform, rather than to perform an import that will leave the object in a situation where Terraform cannot meaningfully manage it.