AWS CDK Custom Resources


cdk-constructs: Stable

Provider Framework

AWS CloudFormation custom resources are extension points to the provisioning engine. When CloudFormation needs to create, update or delete a custom resource, it sends a lifecycle event notification to a custom resource provider. The provider handles the event (e.g. creates a resource) and sends back a response to CloudFormation.

The @aws-cdk/custom-resources.Provider construct is a “mini-framework” for implementing providers for AWS CloudFormation custom resources. The framework offers a high-level API which makes it easier to implement robust and powerful custom resources and includes the following capabilities:

  • Handles responses to AWS CloudFormation and protects against blocked deployments

  • Validates handler return values to help with correct handler implementation

  • Supports asynchronous handlers to enable operations that require a long waiting period for a resource, which can exceed the AWS Lambda timeout

  • Implements default behavior for physical resource IDs.

The following code shows how the Provider construct is used in conjunction with a CustomResource and a user-provided AWS Lambda function which implements the actual handler.

# Example automatically generated without compilation. See
from aws_cdk.core import CustomResource
import aws_cdk.aws_logs as logs
import aws_cdk.aws_iam as iam
import aws_cdk.custom_resources as cr

on_event = lambda_.Function(self, "MyHandler")

my_role = iam.Role(self, "MyRole")

my_provider = cr.Provider(self, "MyProvider",
    is_complete_handler=is_complete, # optional async "waiter"
    log_retention=logs.RetentionDays.ONE_DAY, # default is INFINITE

CustomResource(self, "Resource1", service_token=my_provider.service_token)
CustomResource(self, "Resource2", service_token=my_provider.service_token)

Providers are implemented through AWS Lambda functions that are triggered by the provider framework in response to lifecycle events.

At the minimum, users must define the onEvent handler, which is invoked by the framework for all resource lifecycle events (Create, Update and Delete) and returns a result which is then submitted to CloudFormation.

The following example is a skeleton for a Python implementation of onEvent:

def on_event(event, context):
  request_type = event['RequestType']
  if request_type == 'Create': return on_create(event)
  if request_type == 'Update': return on_update(event)
  if request_type == 'Delete': return on_delete(event)
  raise Exception("Invalid request type: %s" % request_type)

def on_create(event):
  props = event["ResourceProperties"]
  print("create new resource with props %s" % props)

  # add your create code here...
  physical_id = ...

  return { 'PhysicalResourceId': physical_id }

def on_update(event):
  physical_id = event["PhysicalResourceId"]
  props = event["ResourceProperties"]
  print("update resource %s with props %s" % (physical_id, props))
  # ...

def on_delete(event):
  physical_id = event["PhysicalResourceId"]
  print("delete resource %s" % physical_id)
  # ...

Users may also provide an additional handler called isComplete, for cases where the lifecycle operation cannot be completed immediately. The isComplete handler will be retried asynchronously after onEvent until it returns IsComplete: true, or until the total provider timeout has expired.

The following example is a skeleton for a Python implementation of isComplete:

def is_complete(event, context):
  physical_id = event["PhysicalResourceId"]
  request_type = event["RequestType"]

  # check if resource is stable based on request_type
  is_ready = ...

  return { 'IsComplete': is_ready }

Handling Lifecycle Events: onEvent

The user-defined onEvent AWS Lambda function is invoked whenever a resource lifecycle event occurs. The function is expected to handle the event and return a response to the framework that, at least, includes the physical resource ID.

If onEvent returns successfully, the framework will submit a “SUCCESS” response to AWS CloudFormation for this resource operation. If the provider is asynchronous (isCompleteHandler is defined), the framework will only submit a response based on the result of isComplete.

If onEvent throws an error, the framework will submit a “FAILED” response to AWS CloudFormation.

The input event includes the following fields derived from the [Custom Resource Provider Request]:

Field Type Description
RequestType String The type of lifecycle event: Create, Update or Delete.
LogicalResourceId String The template developer-chosen name (logical ID) of the custom resource in the AWS CloudFormation template.
PhysicalResourceId String This field will only be present for Update and Delete events and includes the value returned in PhysicalResourceId of the previous operation.
ResourceProperties JSON This field contains the properties defined in the template for this custom resource.
OldResourceProperties JSON This field will only be present for Update events and contains the resource properties that were declared previous to the update request.
ResourceType String The resource type defined for this custom resource in the template. A provider may handle any number of custom resource types.
RequestId String A unique ID for the request.
StackId String The ARN that identifies the stack that contains the custom resource.

The return value from onEvent must be a JSON object with the following fields:

Field Type Required Description
PhysicalResourceId String No The allocated/assigned physical ID of the resource. If omitted for Create events, the event's RequestId will be used. For Update, the current physical ID will be used. If a different value is returned, CloudFormation will follow with a subsequent Delete for the previous ID (resource replacement). For Delete, it will always return the current physical resource ID, and if the user returns a different one, an error will occur.
Data JSON No Resource attributes, which can later be retrieved through Fn::GetAtt on the custom resource object.
any any No Any other field included in the response will be passed through to isComplete. This can sometimes be useful to pass state between the handlers.

Asynchronous Providers: isComplete

It is not uncommon for the provisioning of resources to be an asynchronous operation, which means that the operation does not immediately finish, and we need to “wait” until the resource stabilizes.

The provider framework makes it easy to implement “waiters” by allowing users to specify an additional AWS Lambda function in isCompleteHandler.

The framework will repeatedly invoke the handler every queryInterval. When isComplete returns with IsComplete: true, the framework will submit a “SUCCESS” response to AWS CloudFormation. If totalTimeout expires and the operation has not yet completed, the framework will submit a “FAILED” response with the message “Operation timed out”.

If an error is thrown, the framework will submit a “FAILED” response to AWS CloudFormation.

The input event to isComplete includes all request fields, combined with all fields returned from onEvent. If PhysicalResourceId has not been explicitly returned from onEvent, it’s value will be calculated based on the heuristics described above.

The return value must be a JSON object with the following fields:

Field Type Required Description
IsComplete Boolean Yes Indicates if the operation has finished or not.
Data JSON No May only be sent if IsComplete is true and includes additional resource attributes. These attributes will be merged with the ones returned from onEvent

Physical Resource IDs

Every resource in CloudFormation has a physical resource ID. When a resource is created, the PhysicalResourceId returned from the Create operation is stored by AWS CloudFormation and assigned to the logical ID defined for this resource in the template. If a Create operation returns without a PhysicalResourceId, the framework will use RequestId as the default. This is sufficient for various cases such as “pseudo-resources” which only query data.

For Update and Delete operations, the resource event will always include the current PhysicalResourceId of the resource.

When an Update operation occurs, the default behavior is to return the current physical resource ID. if the onEvent returns a PhysicalResourceId which is different from the current one, AWS CloudFormation will treat this as a resource replacement, and it will issue a subsequent Delete operation for the old resource.

As a rule of thumb, if your custom resource supports configuring a physical name (e.g. you can specify a BucketName when you define an AWS::S3::Bucket), you must return this name in PhysicalResourceId and make sure to handle replacement properly. The S3File example demonstrates this through the objectKey property.

Handling Provider Framework Error

As mentioned above, if any of the user handlers fail (i.e. throws an exception) or times out (due to their AWS Lambda timing out), the framework will trap these errors and submit a “FAILED” response to AWS CloudFormation, along with the error message.

Since errors can occur in multiple places in the provider (framework, onEvent, isComplete), it is important to know that there could situations where a resource operation fails even though the operation technically succeeded (i.e. isComplete throws an error).

When AWS CloudFormation receives a “FAILED” response, it will attempt to roll back the stack to it’s last state. This has different meanings for different lifecycle events:

  • If a Create event fails, the resource provider framework will automatically ignore the subsequent Delete operation issued by AWS CloudFormation. The framework currently does not support customizing this behavior (see

  • If an Update event fails, CloudFormation will issue an additional Update with the previous properties.

  • If a Delete event fails, CloudFormation will abandon this resource.

Provider Framework Execution Policy

Similarly to any AWS Lambda function, if the user-defined handlers require access to AWS resources, you will have to define these permissions by calling “grant” methods such as myBucket.grantRead(myHandler)), using myHandler.addToRolePolicy or specifying an initialPolicy when defining the function.

Bear in mind that in most cases, a single provider will be used for multiple resource instances. This means that the execution policy of the provider must have the appropriate privileges.

The following example grants the onEvent handler s3:GetObject* permissions to all buckets:

# Example automatically generated without compilation. See
lambda_.Function(self, "OnEventHandler",
    # ...
        iam.PolicyStatement(actions=["s3:GetObject*"], resources=["*"])


Users are responsible to define the timeouts for the AWS Lambda functions for user-defined handlers. It is recommended not to exceed a 14 minutes timeout, since all framework functions are configured to time out after 15 minutes, which is the maximal AWS Lambda timeout.

If your operation takes over 14 minutes, the recommended approach is to implement an asynchronous provider, and then configure the timeouts for the asynchronous retries through the queryInterval and the totalTimeout options.

Provider Framework Examples

This module includes a few examples for custom resource implementations:


Provisions an object in an S3 bucket with textual contents. See the source code for the construct and handler.

The following example will create the file folder/file1.txt inside myBucket with the contents hello!.

# Example automatically generated without compilation. See
S3File(self, "MyFile",
    object_key="folder/file1.txt", # optional

This sample demonstrates the following concepts:

  • Synchronous implementation (isComplete is not defined)

  • Automatically generates the physical name if objectKey is not defined

  • Handles physical name changes

  • Returns resource attributes

  • Handles deletions

  • Implemented in TypeScript


Checks that the textual contents of an S3 object matches a certain value. The check will be retried for 5 minutes as long as the object is not found or the value is different. See the source code for the construct and handler.

The following example defines an S3Assert resource which waits until myfile.txt in myBucket exists and includes the contents foo bar:

# Example automatically generated without compilation. See
S3Assert(self, "AssertMyFile",
    expected_content="foo bar"

This sample demonstrates the following concepts:

  • Asynchronous implementation

  • Non-intrinsic physical IDs

  • Implemented in Python

Custom Resources for AWS APIs

Sometimes a single API call can fill the gap in the CloudFormation coverage. In this case you can use the AwsCustomResource construct. This construct creates a custom resource that can be customized to make specific API calls for the CREATE, UPDATE and DELETE events. Additionally, data returned by the API call can be extracted and used in other constructs/resources (creating a real CloudFormation dependency using Fn::GetAtt under the hood).

The physical id of the custom resource can be specified or derived from the data returned by the API call.

The AwsCustomResource uses the AWS SDK for JavaScript. Services, actions and parameters can be found in the API documentation.

Path to data must be specified using a dot notation, e.g. to get the string value of the Title attribute for the first item returned by dynamodb.query it should be Items.0.Title.S.

To make sure that the newest API calls are available the latest AWS SDK v2 is installed in the Lambda function implementing the custom resource. The installation takes around 60 seconds. If you prefer to optimize for speed, you can disable the installation by setting the installLatestAwsSdk prop to false.

Custom Resource Execution Policy

You must provide the policy property defining the IAM Policy that will be applied to the API calls. The library provides two factory methods to quickly configure this:

  • AwsCustomResourcePolicy.fromSdkCalls - Use this to auto-generate IAM Policy statements based on the configured SDK calls. Note that you will have to either provide specific ARN’s, or explicitly use AwsCustomResourcePolicy.ANY_RESOURCE to allow access to any resource.

  • AwsCustomResourcePolicy.fromStatements - Use this to specify your own custom statements.

The custom resource also implements iam.IGrantable, making it possible to use the grantXxx() methods.

As this custom resource uses a singleton Lambda function, it’s important to note that the function’s role will eventually accumulate the permissions/grants from all resources.

Chained API calls can be achieved by creating dependencies:

# Example automatically generated without compilation. See
aws_custom1 = AwsCustomResource(self, "API1",
        "service": "...",
        "action": "...",
        "physical_resource_id": PhysicalResourceId.of("...")

aws_custom2 = AwsCustomResource(self, "API2",
        "service": "...",
        "action": "...",
        "parameters": {
            "text": aws_custom1.get_response_field("Items.0.text")
        "physical_resource_id": PhysicalResourceId.of("...")

Physical Resource Id Parameter

Some AWS APIs may require passing the physical resource id in as a parameter for doing updates and deletes. You can pass it by using PhysicalResourceIdReference.

# Example automatically generated without compilation. See
aws_custom = AwsCustomResource(self, "...", {
    "on_create": {
        "service": "...",
        "action": "...",
        "parameters": {
            "text": "..."
        "physical_resource_id": PhysicalResourceId.of("...")
    "on_update": {
        "service": "...",
        "action": "...".parameters
}, {
    "text": "...",
    "resource_id": PhysicalResourceIdReference()
}, policy, AwsCustomResourcePolicy.from_sdk_calls(resources=AwsCustomResourcePolicy.ANY_RESOURCE))

Handling Custom Resource Errors

Every error produced by the API call is treated as is and will cause a “FAILED” response to be submitted to CloudFormation. You can ignore some errors by specifying the ignoreErrorCodesMatching property, which accepts a regular expression that is tested against the code property of the response. If matched, a “SUCCESS” response is submitted. Note that in such a case, the call response data and the Data key submitted to CloudFormation would both be an empty JSON object. Since a successful resource provisioning might or might not produce outputs, this presents us with some limitations:

  • PhysicalResourceId.fromResponse - Since the call response data might be empty, we cannot use it to extract the physical id.

  • getResponseField and getResponseFieldReference - Since the Data key is empty, the resource will not have any attributes, and therefore, invoking these functions will result in an error.

In both the cases, you will get a synth time error if you attempt to use it in conjunction with ignoreErrorCodesMatching.

Customizing the Lambda function implementing the custom resource

Use the role, timeout, logRetention and functionName properties to customize the Lambda function implementing the custom resource:

# Example automatically generated without compilation. See
AwsCustomResource(self, "Customized",
    # other props here
    role=my_role, # must be assumable by the `` service principal
    timeout=cdk.Duration.minutes(10), # defaults to 2 minutes
    log_retention=logs.RetentionDays.ONE_WEEK, # defaults to never delete logs

Restricting the output of the Custom Resource

CloudFormation imposes a hard limit of 4096 bytes for custom resources response objects. If your API call returns an object that exceeds this limit, you can restrict the data returned by the custom resource to specific paths in the API response:

# Example automatically generated without compilation. See
AwsCustomResource(stack, "ListObjects",
        "service": "s3",
        "action": "listObjectsV2",
        "parameters": {
            "Bucket": "my-bucket"
        "physical_resource_id": PhysicalResourceId.of("id"),
        "output_paths": ["Contents.0.Key", "Contents.1.Key"]

Note that even if you restrict the output of your custom resource you can still use any path in PhysicalResourceId.fromResponse().

Custom Resource Examples

Verify a domain with SES

# Example automatically generated without compilation. See
verify_domain_identity = AwsCustomResource(self, "VerifyDomainIdentity",
        "service": "SES",
        "action": "verifyDomainIdentity",
        "parameters": {
            "Domain": ""
        "physical_resource_id": PhysicalResourceId.from_response("VerificationToken")

route53.TxtRecord(self, "SESVerificationRecord",

Get the latest version of a secure SSM parameter

# Example automatically generated without compilation. See
get_parameter = AwsCustomResource(self, "GetParameter",
    on_update={# will also be called for a CREATE event
        "service": "SSM",
        "action": "getParameter",
        "parameters": {
            "Name": "my-parameter",
            "WithDecryption": True
        "physical_resource_id": PhysicalResourceId.of(},

# Use the value in another construct with

Associate a PrivateHostedZone with VPC shared from another account

# Example automatically generated without compilation. See
get_parameter = AwsCustomResource(self, "AssociateVPCWithHostedZone",
        "assumed_role_arn": "arn:aws:iam::OTHERACCOUNT:role/CrossAccount/ManageHostedZoneConnections",
        "service": "Route53",
        "action": "associateVPCWithHostedZone",
        "parameters": {
            "HostedZoneId": "hz-123",
            "VPC": {
                "VPCId": "vpc-123",
                "VPCRegion": "region-for-vpc"
        "physical_resource_id": PhysicalResourceId.of("${vpcStack.SharedVpc.VpcId}-${vpcStack.Region}-${PrivateHostedZone.HostedZoneId}")
    # Will ignore any resource and use the assumedRoleArn as resource and 'sts:AssumeRole' for service:action

This module is part of the AWS Cloud Development Kit project.