Resource type handler contract - Extension Development for CloudFormation
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Resource type handler contract

The resource type handler contract specifies the expected and required behavior to which a resource must adhere in each given event handler. It defines a set of specific, unambiguous rules with which create, read, update, delete and list resource handlers must comply. Following the contract will allow customers to interact with all resource types under a uniform set of behaviors and expectations, and prevents creation of unintended or duplicate resources.

A resource implementation MUST pass all resource contract tests to be registered.

Assuming no other concurrent interaction on the resource, the handlers must comply with the following contract.

All terminology in the handler contract requirements adheres to the RFC 2119 specification.

Create handlers

CloudFormation invokes the create handler when the resource is created during a stack create operation.

Input assumptions

The create handler can make the following assumptions about input submitted to it:

  • The input to a create handler MUST be valid against the resource schema.

Output requirements

The create handler must adhere to the following requirements regarding its output:

  • A create handler MUST always return a ProgressEvent object within 60 seconds. For more information, see ProgressEvent Object Schema.

    In every ProgressEvent object, the create handler MUST return a model which conforms to the shape of the resource schema. For more information, see Returned models must conform to the shape of the schema.

    Every model MUST include the primaryIdentifier. The only exception is if the first progress event is FAILED, and the resource hasn't yet been created. In this case, a subsequent read call MUST return NotFound.

  • A create handler MUST NOT return SUCCESS until it has applied all properties included in the create request. For more information, see Update, create, and delete handlers must satisfy desired-state stabilization.

    • A create handler MUST return IN_PROGRESS if it hasn't yet reached the desired-state.

      A create handler SHOULD return a model containing all properties set so far and nothing more during each IN_PROGRESS event.

    • A create handler MUST return FAILED progress event if it can't reach the desired-state within the timeout specified in the resource schema.

      The progress event MUST return an error message and the most applicable error code. For more information, see Handler error codes.

    • A create handler MAY return SUCCESS once it reaches the desired-state.

      Once the desired state has been reached, a create handler MAY perform runtime-state stabilization. For more information, see Update and create handlers should satisfy runtime-state stabilization.

      When the create handler returns SUCCESS, it MUST return a ProgressEvent object containing a model that satisfies the following requirements:

      • All properties specified in the create request MUST be present in the model returned, and they MUST match exactly, with the exception of properties defined as writeOnlyProperties in the resource schema.

      • The model MUST contain all properties that have values, including any properties that have default values, and any readOnlyProperties as defined in the resource schema.

      • The model MUST NOT return any properties that are null or don't have values.

  • After a create operation returns SUCCESS, a subsequent read request MUST succeed when passed in the primaryIdentifier or any additionalIdentifiers associated with the provisioned resource instance.

  • After a create operation returns SUCCESS, a subsequent list operation MUST return the primaryIdentifier associated with the provisioned resource instance.

    If the list operation is paginated, the entire list operation is defined as all list requests until the nextToken is null.

  • A create handler MUST be idempotent. A create handler MUST NOT create multiple resources given the same idempotency token.

  • A create handler MUST return FAILED with an AlreadyExists error code if the resource already existed before the create request.

Update handlers

CloudFormation invokes the update handler when the resource is updated during an update operation.

Input assumptions

The update handler can make the following assumptions about input submitted to it:

  • The input to an update handler MUST be valid against the resource schema.

  • Any createOnlyProperties specified in update handler input MUST NOT be different from their previous state.

  • The input to an update handler MUST contain either the primaryIdentifier or an additionalIdentifier.

Output requirements

The update handler must adhere to the following requirements:

  • An update handler MUST always return a ProgressEvent object within 60 seconds. For more information, see ProgressEvent Object Schema.

    In every ProgressEvent object, the update handler MUST return a model which conforms to the shape of the resource schema. For more information, see Returned models must conform to the shape of the schema.

    Every model MUST include the primaryIdentifier.

    The primaryIdentifier returned in every progress event must match the primaryIdentifier passed into the request.

  • An update handler MUST NOT return SUCCESS until it has applied all properties included in the update request. For more information, see Update, create, and delete handlers must satisfy desired-state stabilization.

    • An update handler MUST return IN_PROGRESS if it hasn't yet reached the desired-state.

      An update handler SHOULD return a model containing all properties set so far and nothing more during each IN_PROGRESS event.

    • An update handler MUST return FAILED progress event if it can't reach the desired-state within the timeout specified in the resource schema.

      The progress event MUST return an error message and the most applicable error code. For more information, see Handler error codes.

    • An update handler MAY return SUCCESS once it reaches the desired-state.

      Once the desired state has been reached, an update handler MAY perform runtime-state stabilization. For more information, see Update and create handlers should satisfy runtime-state stabilization.

      When the update handler returns SUCCESS, it MUST return a ProgressEvent object containing a model that satisfies the following requirements:

      • All properties specified in the update request MUST be present in the model returned, and they MUST match exactly, with the exception of properties defined as writeOnlyProperties in the resource schema.

      • The model MUST contain all properties that have values, including any properties that have default values, and any readOnlyProperties as defined in the resource schema.

      • The model MUST NOT return any properties that are null or don't have values.

    All list or collection properties MUST be applied in full. The successful outcome MUST be replacement of the previous properties, if any.

  • An update handler MUST return FAILED with a NotFound error code if the resource didn't exist before the update request.

  • An update handler MUST NOT create a new physical resource.

Delete handlers

CloudFormation invokes the delete handler when the resource, or entire stack, is deleted during a stack delete operation.

Input assumptions

The delete handler can make the following assumptions about input submitted to it:

  • The input to a delete handler MUST contain either the primaryIdentifier or an additionalIdentifier. Any other properties MAY NOT be included in the request.

Output requirements

The delete handler must adhere to the following requirements:

  • A delete handler MUST always return a ProgressEvent object within 60 seconds. For more information, see ProgressEvent Object Schema.

  • A delete handler MUST NOT return SUCCESS until the resource has reached the desired state for deletion. For more information, see Update, create, and delete handlers must satisfy desired-state stabilization.

    • A delete handler MUST return IN_PROGRESS if it hasn't yet reached the desired state.

    • A delete handler MUST return FAILED progress event if it can't reach the desired-state within the timeout specified in the resource schema.

      The progress event MUST return an error message and the most applicable error code. For more information, see Handler error codes.

    • A delete handler MUST return SUCCESS once it reaches the desired state. (This is because there is no runtime-state stabilization for delete requests.)

      When the delete handler returns SUCCESS, the ProgressEvent object MUST NOT contain a model.

  • A delete handler MUST return FAILED with a NotFound error code if the resource didn't exist before the delete request.

  • Once a delete operation successfully completes, any subsequent update, delete, or read request for the deleted resource instance MUST return FAILED with a NotFound error code.

  • Once a delete operation successfully completes, any subsequent list operation MUST NOT return the primaryIdentifier associated with the deleted resource instance.

    If the list operation is paginated, the 'list operation' is defined as all list calls until the nextToken is null.

  • Once a delete operation successfully completes, a subsequent create request with the same primaryIdentifier or additionalIdentifiers MUST NOT return FAILED with an AlreadyExists error code.

  • Once a delete operation successfully completes, the resource SHOULD NOT be billable to the client.

Read handlers

CloudFormation invokes the read handler when detailed information about the resource needed during a stack update operation.

Input assumptions

The read handler can make the following assumptions about input submitted to it:

  • The input to a read handler MUST contain either the primaryIdentifier or an additionalIdentifier. Any other properties MAY NOT be included in the request.

Output requirements

The read handler must adhere to the following requirements regarding its output:

  • A read handler MUST always return a ProgressEvent object within 30 seconds. For more information, see ProgressEvent Object Schema.

    A read handler MUST always return a status of SUCCESS or FAILED; it MUST NOT return a status of IN_PROGRESS.

  • A read handler MUST return a model representation that conforms to the shape of the resource schema.

    • The model MUST contain all properties that have values, including any properties that have default values and any readOnlyProperties as defined in the resource schema.

    • The model MUST NOT return any properties that are null or don't have values.

  • A read handler MUST return FAILED with a NotFound error code if the resource doesn't exist.

List handlers

CloudFormation invokes the list handler when summary information about multiple resources of this resource type is required.

  • A list handler MUST always return a ProgressEvent object within 30 seconds. For more information, see ProgressEvent Object Schema.

    A list handler MUST always return a status of SUCCESS or FAILED; it MUST NOT return a status of IN_PROGRESS.

  • A list handler MUST return an array of primary identifiers.

    When passed in a read request, each primaryIdentifier MUST NOT return FAILED with NotFound error code.

  • A list request MUST support pagination by returning a NextToken.

    The NextToken returned MUST be able to be used in a subsequent list request to retrieve the next set of results from the service.

    The NextToken MUST be null when all results have been returned.

  • A list request MUST return an empty array if there are no resources found.

  • A list handler MAY accept a set of properties conforming to the shape of the resource schema as filter criteria.

    The filter should use AND(&) when multiple properties are passed in.

Additional requirements

The following requirements also apply to resource handlers.

Returned models must conform to the shape of the schema

A model returned in a ProgressEvent object MUST always conform to the shape of the resource schema. This means that each property that's returned MUST adhere to its own individual restrictions: correct data type, regex, length, etc. However, the model returned MAY NOT contain all properties defined as required in the json-schema.

More specifically, contract tests validate models based on json-schema Validation keywords.

  • ALL Validation Keywords for the following MUST be observed:

    • Any Instance Type (Section 6.1)

    • Numeric Instances (Section 6.2)

    • Strings (Section 6.3)

    • Arrays (Section 6.4)

  • All Validation Keywords for Objects (Section 6.5) MUST be observed EXCEPT for:

    • required (Section 6.5.3)

    • dependencies (Section 6.5.7)

    • propertyNames (Section 6.5.8)

  • Contract tests won't validate Validation Keywords for:

    • Applying Subschemas Conditionally (Section 6.6)

    • Applying Subschemas With Boolean Logic (Section 6.7)

Update, create, and delete handlers must satisfy desired-state stabilization

Stabilization is the process of waiting for a resource to be in a particular state. Note that reaching the desired-state is mandatory for all handlers before returning SUCCESS.

Create and update handlers

For Create and Update handlers, desired-state stabilization is satisfied when all properties specified in the request are applied as requested. This is verified by calling the Read handler.

In many cases, the desired-state is reached immediately upon completion of a Create/Update API call. However, in some cases, multiple API calls and or wait periods may be required in order to reach this state.

Eventual consistency in desired-state stabilization

Eventual consistency means that the result of an API command you run might not be immediately visible to all subsequent commands you run. Handling API eventual consistency is required as part of desired-state stabilization. This is because a subsequent Read call might fail with a NotFound error code.

Amazon EC2 resources are a great example of this. For more information, see Eventual Consistency in the Amazon Elastic Compute Cloud API Reference.

Examples of desired-state stabilization

For a simple example of desired-state stabilization, consider the implementation of the create handler for the AWS::Logs::MetricFilter resource: immediately after the handler code completes the call to the PutMetricFilter method, the AWS::Logs::MetricFilter has achieved its desired state. You can examine the code for this resource in its open-source repository at github.com/aws-cloudformation/aws-cloudformation-resource-providers-logs.

A more complex example is the implementation of the update handler for the AWS::Kinesis::Stream resource. The update handler must make multiple API calls during an update, including AddTagsToStream or RemoveTagsFromStream, UpdateShardCount, IncreaseRetentionPeriod or DecreaseRetentionPeriod, and StartStreamEncryption or StopStreamEncryption. Meanwhile, each API call will set the StreamStatus to UPDATING, during which time other API operations can't be performed or the API will throw a ResourceInUseException. Therefore, to reach the desired state, the handler will need to wait for the StreamStatus to become ACTIVE in between each API operation.

Delete handlers

Usually, the definition of deleted is obvious. A delete API operation will result in the resource being purged from the database, and the resource is no longer describable to the user.

However, sometimes, a deletion will result in the resource leaving an audit trail, in which the resource can still be described by service API operations, but can no longer be interacted with by the user. For example, when you delete a CloudFormation stack, it's assigned a status of DELETE_COMPLETE, but it can still be returned from a DescribeStacks API call. For resources like this, the desired-state for deletion is when the resource has reached a terminal, inoperable, and irrecoverable state. If the resource can continue to be mutated by the user through another API call, then it isn't deleted, it's updated.

There is no difference between desired-state stabilization and runtime-state stabilization for a delete handler. By definition, once a resource has reached the desired-state for deletion, a subsequent read call MUST return FAILED with a NotFound error code, and a subsequent create call with the same primaryIdentifier or additionalIdentifiers MUST NOT return FAILED with an AlreadyExists error code. Additional restrictions are defined in the contract above.

So in the case of a CloudFormation stack, a read handler MUST return FAILED with a NotFound error code if the stack is DELETE_COMPLETE, even though it's audit trail can still be accessed by the DescribeStacks API.

Update and create handlers should satisfy runtime-state stabilization

Runtime-state stabilization is a process of waiting for the resource to be "ready" to use. Generally, runtime-state stabilization is done by continually describing the resource until it reaches a particular state, though it can take many forms.

Runtime-state stabilization can mean different things for different resources, but the following are common requirements:

  • Additional mutating API calls can be made on the resource

    Some resources can't be modified while they're in a particular state.

  • Dependent resources can consume the resource

    There may be other resources which need to consume the resource in some way, but can't until it is in a particular state.

  • Users can interact with the resource

    Customers may not be able to use the resource until it is in a particular status. This usually overlaps with the dependent resources requirement, although there could be different qualifications, depending on the resources.

While desired-state stabilization is mandatory, runtime-state stabilization is optional but encouraged. Users have come to expect that once a resource is COMPLETE, they will be able to use it.

Examples of run-time stabilization

For a simple example of run-time stabilization, consider the implementation of the create handler for the AWS::KinesisFirehose::DeliveryStream resource. The create handler invokes only a single API, CreateDeliveryStream, in order for the resource to reach its desired state. Immediately after this API call is made, a read request will return the correct desired state. However, the resource still has not reached run-time stabilization because it can't be used by the customer or downstream resources until the DeliveryStreamStatus is ACTIVE.

For a more complex example, consider the implementation of the update handler for the AWS::Kinesis::Stream resource once again. Once the update handler has made its final call, to StartStreamEncryption or StopStreamEncryption as described in Examples of desired-state stabilization, the resource has reached its desired state. However, like the other API calls on the Kinesis resource, the StreamStatus will again be set to UPDATING. During this period, it has reached its desired state, and customers can even continue using the stream. But it hasn't yet achieved runtime-stabilization, because additional API calls cannot be made on the resource until the StreamStatus gets set to ACTIVE.

Handlers must not leak resources

Resource leaking refers to when a handler loses track of the existence of a resource. This happens most often in the following cases:

  • A create handler isn't idempotent. Re-invoking the handler with the same idempotencyToken will cause another resource to be created, and the handler is only tracking a single resource.

  • A create handler creates the resource, but is unable to communicate an identifier for that resource back to CloudFormation. A subsequent delete call doesn't have enough information to delete the resource.

  • A bug in the delete handler causes the resource to not actually be deleted, but the delete handler reports that the resource was successfully deleted.