AWS Cloud Development Kit Core Library


cfn-resources: Stable

cdk-constructs: Stable

This library includes the basic building blocks of the AWS Cloud Development Kit (AWS CDK). It defines the core classes that are used in the rest of the AWS Construct Library.

See the AWS CDK Developer Guide for information of most of the capabilities of this library. The rest of this README will only cover topics not already covered in the Developer Guide.

Stacks and Stages

A Stack is the smallest physical unit of deployment, and maps directly onto a CloudFormation Stack. You define a Stack by defining a subclass of Stack – let’s call it MyStack – and instantiating the constructs that make up your application in MyStack’s constructor. You then instantiate this stack one or more times to define different instances of your application. For example, you can instantiate it once using few and cheap EC2 instances for testing, and once again using more and bigger EC2 instances for production.

When your application grows, you may decide that it makes more sense to split it out across multiple Stack classes. This can happen for a number of reasons:

  • You could be starting to reach the maximum number of resources allowed in a single stack (this is currently 500).

  • You could decide you want to separate out stateful resources and stateless resources into separate stacks, so that it becomes easy to tear down and recreate the stacks that don’t have stateful resources.

  • There could be a single stack with resources (like a VPC) that are shared between multiple instances of other stacks containing your applications.

As soon as your conceptual application starts to encompass multiple stacks, it is convenient to wrap them in another construct that represents your logical application. You can then treat that new unit the same way you used to be able to treat a single stack: by instantiating it multiple times for different instances of your application.

You can define a custom subclass of Construct, holding one or more Stacks, to represent a single logical instance of your application.

As a final note: Stacks are not a unit of reuse. They describe physical deployment layouts, and as such are best left to application builders to organize their deployments with. If you want to vend a reusable construct, define it as a subclasses of Construct: the consumers of your construct will decide where to place it in their own stacks.

Nested Stacks

Nested stacks are stacks created as part of other stacks. You create a nested stack within another stack by using the NestedStack construct.

As your infrastructure grows, common patterns can emerge in which you declare the same components in multiple templates. You can separate out these common components and create dedicated templates for them. Then use the resource in your template to reference other templates, creating nested stacks.

For example, assume that you have a load balancer configuration that you use for most of your stacks. Instead of copying and pasting the same configurations into your templates, you can create a dedicated template for the load balancer. Then, you just use the resource to reference that template from within other templates.

The following example will define a single top-level stack that contains two nested stacks: each one with a single Amazon S3 bucket:

# Example automatically generated. See
class MyNestedStack(cfn.NestedStack):
    def __init__(self, scope, id, *, parameters=None, timeout=None, notifications=None):
        super().__init__(scope, id, parameters=parameters, timeout=timeout, notifications=notifications)

        s3.Bucket(self, "NestedBucket")

class MyParentStack(Stack):
    def __init__(self, scope, id, *, description=None, env=None, stackName=None, tags=None, synthesizer=None, terminationProtection=None, analyticsReporting=None):
        super().__init__(scope, id, description=description, env=env, stackName=stackName, tags=tags, synthesizer=synthesizer, terminationProtection=terminationProtection, analyticsReporting=analyticsReporting)

        MyNestedStack(self, "Nested1")
        MyNestedStack(self, "Nested2")

Resources references across nested/parent boundaries (even with multiple levels of nesting) will be wired by the AWS CDK through CloudFormation parameters and outputs. When a resource from a parent stack is referenced by a nested stack, a CloudFormation parameter will automatically be added to the nested stack and assigned from the parent; when a resource from a nested stack is referenced by a parent stack, a CloudFormation output will be automatically be added to the nested stack and referenced using Fn::GetAtt "Outputs.Xxx" from the parent.

Nested stacks also support the use of Docker image and file assets.

Accessing resources in a different stack

You can access resources in a different stack, as long as they are in the same account and AWS Region. The following example defines the stack stack1, which defines an Amazon S3 bucket. Then it defines a second stack, stack2, which takes the bucket from stack1 as a constructor property.

# Example automatically generated without compilation. See
prod = {"account": "123456789012", "region": "us-east-1"}

stack1 = StackThatProvidesABucket(app, "Stack1", env=prod)

# stack2 will take a property { bucket: IBucket }
stack2 = StackThatExpectsABucket(app, "Stack2",

If the AWS CDK determines that the resource is in the same account and Region, but in a different stack, it automatically synthesizes AWS CloudFormation Exports in the producing stack and an Fn::ImportValue in the consuming stack to transfer that information from one stack to the other.

Removing automatic cross-stack references

The automatic references created by CDK when you use resources across stacks are convenient, but may block your deployments if you want to remove the resources that are referenced in this way. You will see an error like:

Export Stack1:ExportsOutputFnGetAtt-****** cannot be deleted as it is in use by Stack1

Let’s say there is a Bucket in the stack1, and the stack2 references its bucket.bucketName. You now want to remove the bucket and run into the error above.

It’s not safe to remove stack1.bucket while stack2 is still using it, so unblocking yourself from this is a two-step process. This is how it works:

DEPLOYMENT 1: break the relationship

  • Make sure stack2 no longer references bucket.bucketName (maybe the consumer stack now uses its own bucket, or it writes to an AWS DynamoDB table, or maybe you just remove the Lambda Function altogether).

  • In the stack1 class, call this.exportValue(this.bucket.bucketName). This will make sure the CloudFormation Export continues to exist while the relationship between the two stacks is being broken.

  • Deploy (this will effectively only change the stack2, but it’s safe to deploy both).

DEPLOYMENT 2: remove the resource

  • You are now free to remove the bucket resource from stack1.

  • Don’t forget to remove the exportValue() call as well.

  • Deploy again (this time only the stack1 will be changed – the bucket will be deleted).


To make specifications of time intervals unambiguous, a single class called Duration is used throughout the AWS Construct Library by all constructs that that take a time interval as a parameter (be it for a timeout, a rate, or something else).

An instance of Duration is constructed by using one of the static factory methods on it:

# Example automatically generated. See
Duration.seconds(300)# 5 minutes
Duration.minutes(5)# 5 minutes
Duration.hours(1)# 1 hour
Duration.days(7)# 7 days

Size (Digital Information Quantity)

To make specification of digital storage quantities unambiguous, a class called Size is available.

An instance of Size is initialized through one of its static factory methods:

# Example automatically generated. See
Size.kibibytes(200)# 200 KiB
Size.mebibytes(5)# 5 MiB
Size.gibibytes(40)# 40 GiB
Size.tebibytes(200)# 200 TiB

Instances of Size created with one of the units can be converted into others. By default, conversion to a higher unit will fail if the conversion does not produce a whole number. This can be overridden by unsetting integral property.

# Example automatically generated. See
Size.mebibytes(2).to_kibibytes()# yields 2048


To help avoid accidental storage of secrets as plain text, we use the SecretValue type to represent secrets. Any construct that takes a value that should be a secret (such as a password or an access key) will take a parameter of type SecretValue.

The best practice is to store secrets in AWS Secrets Manager and reference them using SecretValue.secretsManager:

# Example automatically generated. See
secret = SecretValue.secrets_manager("secretId",
    json_field="password", # optional: key of a JSON field to retrieve (defaults to all content),
    version_id="id", # optional: id of the version (default AWSCURRENT)

Using AWS Secrets Manager is the recommended way to reference secrets in a CDK app. SecretValue also supports the following secret sources:

  • SecretValue.plainText(secret): stores the secret as plain text in your app and the resulting template (not recommended).

  • SecretValue.ssmSecure(param, version): refers to a secret stored as a SecureString in the SSM Parameter Store.

  • SecretValue.cfnParameter(param): refers to a secret passed through a CloudFormation parameter (must have NoEcho: true).

  • SecretValue.cfnDynamicReference(dynref): refers to a secret described by a CloudFormation dynamic reference (used by ssmSecure and secretsManager).

ARN manipulation

Sometimes you will need to put together or pick apart Amazon Resource Names (ARNs). The functions stack.formatArn() and stack.parseArn() exist for this purpose.

formatArn() can be used to build an ARN from components. It will automatically use the region and account of the stack you’re calling it on:

# Example automatically generated. See
# Builds "arn:<PARTITION>:lambda:<REGION>:<ACCOUNT>:function:MyFunction"

parseArn() can be used to get a single component from an ARN. parseArn() will correctly deal with both literal ARNs and deploy-time values (tokens), but in case of a deploy-time value be aware that the result will be another deploy-time value which cannot be inspected in the CDK application.

# Example automatically generated. See
# Extracts the function name out of an AWS Lambda Function ARN
arn_components = stack.parse_arn(arn, ":")
function_name = arn_components.resource_name

Note that depending on the service, the resource separator can be either : or /, and the resource name can be either the 6th or 7th component in the ARN. When using these functions, you will need to know the format of the ARN you are dealing with.

For an exhaustive list of ARN formats used in AWS, see AWS ARNs and Namespaces in the AWS General Reference.


Construct Dependencies

Sometimes AWS resources depend on other resources, and the creation of one resource must be completed before the next one can be started.

In general, CloudFormation will correctly infer the dependency relationship between resources based on the property values that are used. In the cases where it doesn’t, the AWS Construct Library will add the dependency relationship for you.

If you need to add an ordering dependency that is not automatically inferred, you do so by adding a dependency relationship using constructA.node.addDependency(constructB). This will add a dependency relationship between all resources in the scope of constructA and all resources in the scope of constructB.

If you want a single object to represent a set of constructs that are not necessarily in the same scope, you can use a ConcreteDependable. The following creates a single object that represents a dependency on two constructs, constructB and constructC:

# Example automatically generated. See
# Declare the dependable object
b_and_c = ConcreteDependable()

# Take the dependency

Stack Dependencies

Two different stack instances can have a dependency on one another. This happens when an resource from one stack is referenced in another stack. In that case, CDK records the cross-stack referencing of resources, automatically produces the right CloudFormation primitives, and adds a dependency between the two stacks. You can also manually add a dependency between two stacks by using the stackA.addDependency(stackB) method.

A stack dependency has the following implications:

  • Cyclic dependencies are not allowed, so if stackA is using resources from stackB, the reverse is not possible anymore.

  • Stacks with dependencies between them are treated specially by the CDK toolkit:

    • If stackA depends on stackB, running cdk deploy stackA will also automatically deploy stackB.

    • stackB’s deployment will be performed before stackA’s deployment.

Custom Resources

Custom Resources are CloudFormation resources that are implemented by arbitrary user code. They can do arbitrary lookups or modifications during a CloudFormation deployment.

To define a custom resource, use the CustomResource construct:

# Example automatically generated. See
CustomResource(self, "MyMagicalResource",
    resource_type="Custom::MyCustomResource", # must start with 'Custom::'

    # the resource properties
        "Property1": "foo",
        "Property2": "bar"

    # the ARN of the provider (SNS/Lambda) which handles
    # CREATE, UPDATE or DELETE events for this resource type
    # see next section for details

Custom Resource Providers

Custom resources are backed by a custom resource provider which can be implemented in one of the following ways. The following table compares the various provider types (ordered from low-level to high-level):

Provider Compute Type Error Handling Submit to CloudFormation Max Timeout Language Footprint
sns.Topic Self-managed Manual Manual Unlimited Any Depends
lambda.Function AWS Lambda Manual Manual 15min Any Small
core.CustomResourceProvider Lambda Auto Auto 15min Node.js Small
custom-resources.Provider Lambda Auto Auto Unlimited Async Any Large


  • Compute type: which type of compute can is used to execute the handler.

  • Error Handling: whether errors thrown by handler code are automatically trapped and a FAILED response is submitted to CloudFormation. If this is “Manual”, developers must take care of trapping errors. Otherwise, events could cause stacks to hang.

  • Submit to CloudFormation: whether the framework takes care of submitting SUCCESS/FAILED responses to CloudFormation through the event’s response URL.

  • Max Timeout: maximum allows/possible timeout.

  • Language: which programming languages can be used to implement handlers.

  • Footprint: how many resources are used by the provider framework itself.


When defining resources for a custom resource provider, you will likely want to define them as a stack singleton so that only a single instance of the provider is created in your stack and which is used by all custom resources of that type.

Here is a basic pattern for defining stack singletons in the CDK. The following examples ensures that only a single SNS topic is defined:

# Example automatically generated. See
def get_or_create(self, scope):
    stack = Stack.of(scope)
    uniqueid = "GloballyUniqueIdForSingleton"return stack.node.try_find_child(uniqueid) ?? sns.Topic(stack, uniqueid)

Amazon SNS Topic

Every time a resource event occurs (CREATE/UPDATE/DELETE), an SNS notification is sent to the SNS topic. Users must process these notifications (e.g. through a fleet of worker hosts) and submit success/failure responses to the CloudFormation service.

Set serviceToken to topic.topicArn in order to use this provider:

# Example automatically generated. See
topic = sns.Topic(self, "MyProvider")

CustomResource(self, "MyResource",

AWS Lambda Function

An AWS lambda function is called directly by CloudFormation for all resource events. The handler must take care of explicitly submitting a success/failure response to the CloudFormation service and handle various error cases.

Set serviceToken to lambda.functionArn to use this provider:

# Example automatically generated. See
fn = lambda_.Function(self, "MyProvider", function_props)

CustomResource(self, "MyResource",

The core.CustomResourceProvider class

The class @aws-cdk/core.CustomResourceProvider offers a basic low-level framework designed to implement simple and slim custom resource providers. It currently only supports Node.js-based user handlers, and it does not have support for asynchronous waiting (handler cannot exceed the 15min lambda timeout).

The provider has a built-in singleton method which uses the resource type as a stack-unique identifier and returns the service token:

# Example automatically generated. See
service_token = CustomResourceProvider.get_or_create(self, "Custom::MyCustomResourceType",

CustomResource(self, "MyResource",

The directory (my-handler in the above example) must include an index.js file. It cannot import external dependencies or files outside this directory. It must export an async function named handler. This function accepts the CloudFormation resource event object and returns an object with the following structure:

exports.handler = async function(event) {
  const id = event.PhysicalResourceId; // only for "Update" and "Delete"
  const props = event.ResourceProperties;
  const oldProps = event.OldResourceProperties; // only for "Update"s

  switch (event.RequestType) {
    case "Create":
      // ...

    case "Update":
      // ...

      // if an error is thrown, a FAILED response will be submitted to CFN
      throw new Error('Failed!');

    case "Delete":
      // ...

  return {
    // (optional) the value resolved from `resource.ref`
    // defaults to "event.PhysicalResourceId" or "event.RequestId"
    PhysicalResourceId: "REF",

    // (optional) calling `resource.getAtt("Att1")` on the custom resource in the CDK app
    // will return the value "BAR".
    Data: {
      Att1: "BAR",
      Att2: "BAZ"

    // (optional) user-visible message
    Reason: "User-visible message",

    // (optional) hides values from the console
    NoEcho: true

Here is an complete example of a custom resource that summarizes two numbers:


exports.handler = async (e) => {
  return {
    Data: {
      Result: e.ResourceProperties.lhs + e.ResourceProperties.rhs,


# Example automatically generated. See
from aws_cdk.core import Construct, CustomResource, CustomResourceProvider, CustomResourceProviderRuntime, Token

class Sum(Construct):

    def __init__(self, scope, id, *, lhs, rhs):
        super().__init__(scope, id)

        resource_type = "Custom::Sum"
        service_token = CustomResourceProvider.get_or_create(self, resource_type,

        resource = CustomResource(self, "Resource",
                "lhs": lhs,
                "rhs": rhs

        self.result = Token.as_number(resource.get_att("Result"))

Usage will look like this:

# Example automatically generated. See
sum = Sum(self, "MySum", lhs=40, rhs=2)
CfnOutput(self, "Result", value=Token.as_string(sum.result))

To access the ARN of the provider’s AWS Lambda function role, use the getOrCreateProvider() built-in singleton method:

# Example automatically generated. See
provider = CustomResourceProvider.get_or_create_provider(self, "Custom::MyCustomResourceType",

role_arn = provider.role_arn

This role ARN can then be used in resource-based IAM policies.

The Custom Resource Provider Framework

The @aws-cdk/custom-resources module includes an advanced framework for implementing custom resource providers.

Handlers are implemented as AWS Lambda functions, which means that they can be implemented in any Lambda-supported runtime. Furthermore, this provider has an asynchronous mode, which means that users can provide an isComplete lambda function which is called periodically until the operation is complete. This allows implementing providers that can take up to two hours to stabilize.

Set serviceToken to provider.serviceToken to use this type of provider:

# Example automatically generated. See
provider = customresources.Provider(self, "MyProvider",

CustomResource(self, "MyResource",

See the documentation for more details.

AWS CloudFormation features

A CDK stack synthesizes to an AWS CloudFormation Template. This section explains how this module allows users to access low-level CloudFormation features when needed.

Stack Outputs

CloudFormation stack outputs and exports are created using the CfnOutput class:

# Example automatically generated. See
CfnOutput(self, "OutputName",
    description="The name of an S3 bucket", # Optional


CloudFormation templates support the use of Parameters to customize a template. They enable CloudFormation users to input custom values to a template each time a stack is created or updated. While the CDK design philosophy favors using build-time parameterization, users may need to use CloudFormation in a number of cases (for example, when migrating an existing stack to the AWS CDK).

Template parameters can be added to a stack by using the CfnParameter class:

# Example automatically generated. See
CfnParameter(self, "MyParameter",

The value of parameters can then be obtained using one of the value methods. As parameters are only resolved at deployment time, the values obtained are placeholder tokens for the real value (Token.isUnresolved() would return true for those):

# Example automatically generated. See
param = CfnParameter(self, "ParameterName")

# If the parameter is a String

# If the parameter is a Number

# If the parameter is a List

Pseudo Parameters

CloudFormation supports a number of pseudo parameters, which resolve to useful values at deployment time. CloudFormation pseudo parameters can be obtained from static members of the Aws class.

It is generally recommended to access pseudo parameters from the scope’s stack instead, which guarantees the values produced are qualifying the designated stack, which is essential in cases where resources are shared cross-stack:

# Example automatically generated. See
# "this" is the current construct
stack = Stack.of(self)

stack.account# Returns the AWS::AccountId for this stack (or the literal value if known)
stack.region# Returns the AWS::Region for this stack (or the literal value if known)

Resource Options

CloudFormation resources can also specify resource attributes. The CfnResource class allows accessing those through the cfnOptions property:

# Example automatically generated. See
raw_bucket = s3.CfnBucket(self, "Bucket")
# -or-
raw_bucket_alt = my_bucket.node.default_child

# then
raw_bucket.cfn_options.condition = CfnCondition(self, "EnableBucket")
raw_bucket.cfn_options.metadata = {
    "metadata_key": "MetadataValue"

Resource dependencies (the DependsOn attribute) is modified using the cfnResource.addDependsOn method:

# Example automatically generated. See
resource_a = CfnResource(self, "ResourceA", resource_props)
resource_b = CfnResource(self, "ResourceB", resource_props)


Intrinsic Functions and Condition Expressions

CloudFormation supports intrinsic functions. These functions can be accessed from the Fn class, which provides type-safe methods for each intrinsic function as well as condition expressions:

# Example automatically generated. See
# To use Fn::Base64

# To compose condition expressions:
environment_parameter = CfnParameter(self, "Environment")
    # The "Environment" CloudFormation template parameter evaluates to "Production"
    Fn.condition_equals("Production", environment_parameter),
    # The AWS::Region pseudo-parameter value is NOT equal to "us-east-1"
    Fn.condition_not(Fn.condition_equals("us-east-1", Aws.REGION)))

When working with deploy-time values (those for which Token.isUnresolved returns true), idiomatic conditionals from the programming language cannot be used (the value will not be known until deployment time). When conditional logic needs to be expressed with un-resolved values, it is necessary to use CloudFormation conditions by means of the CfnCondition class:

# Example automatically generated. See
environment_parameter = CfnParameter(self, "Environment")
is_prod = CfnCondition(self, "IsProduction",
    expression=Fn.condition_equals("Production", environment_parameter)

# Configuration value that is a different string based on IsProduction
stage = Fn.condition_if(is_prod.logical_id, "Beta", "Prod").to_string()

# Make Bucket creation condition to IsProduction by accessing
# and overriding the CloudFormation resource
bucket = s3.Bucket(self, "Bucket")
cfn_bucket = my_bucket.node.default_child
cfn_bucket.cfn_options.condition = is_prod


CloudFormation mappings are created and queried using the CfnMappings class:

# Example automatically generated. See
region_table = CfnMapping(self, "RegionTable",
        "region_name": {
            "us-east-1": "US East (N. Virginia)",
            "us-east-2": "US East (Ohio)"

region_table.find_in_map("regionName", Aws.REGION)

This will yield the following template:

      us-east-1: US East (N. Virginia)
      us-east-2: US East (Ohio)

Dynamic References

CloudFormation supports dynamically resolving values for SSM parameters (including secure strings) and Secrets Manager. Encoding such references is done using the CfnDynamicReference class:

# Example automatically generated. See
CfnDynamicReference(CfnDynamicReferenceService.SECRETS_MANAGER, "secret-id:secret-string:json-key:version-stage:version-id")

Template Options & Transform

CloudFormation templates support a number of options, including which Macros or Transforms to use when deploying the stack. Those can be configured using the stack.templateOptions property:

# Example automatically generated. See
stack = Stack(app, "StackName")

stack.template_options.description = "This will appear in the AWS console"
stack.template_options.transforms = ["AWS::Serverless-2016-10-31"]
stack.template_options.metadata = {
    "metadata_key": "MetadataValue"

Emitting Raw Resources

The CfnResource class allows emitting arbitrary entries in the Resources section of the CloudFormation template.

# Example automatically generated. See
CfnResource(self, "ResourceId",
        "BucketName": "bucket-name"

As for any other resource, the logical ID in the CloudFormation template will be generated by the AWS CDK, but the type and properties will be copied verbatim in the synthesized template.

Including raw CloudFormation template fragments

When migrating a CloudFormation stack to the AWS CDK, it can be useful to include fragments of an existing template verbatim in the synthesized template. This can be achieved using the CfnInclude class.

# Example automatically generated. See
CfnInclude(self, "ID",
        "Resources": {
            "Bucket": {
                "Type": "AWS::S3::Bucket",
                "Properties": {
                    "BucketName": "my-shiny-bucket"

Termination Protection

You can prevent a stack from being accidentally deleted by enabling termination protection on the stack. If a user attempts to delete a stack with termination protection enabled, the deletion fails and the stack–including its status–remains unchanged. Enabling or disabling termination protection on a stack sets it for any nested stacks belonging to that stack as well. You can enable termination protection on a stack by setting the terminationProtection prop to true.

# Example automatically generated. See
stack = Stack(app, "StackName",

By default, termination protection is disabled.


CfnJson allows you to postpone the resolution of a JSON blob from deployment-time. This is useful in cases where the CloudFormation JSON template cannot express a certain value.

A common example is to use CfnJson in order to render a JSON map which needs to use intrinsic functions in keys. Since JSON map keys must be strings, it is impossible to use intrinsics in keys and CfnJson can help.

The following example defines an IAM role which can only be assumed by principals that are tagged with a specific tag.

# Example automatically generated. See
tag_param = CfnParameter(self, "TagName")

string_equals = CfnJson(self, "ConditionJson",
        f"aws:PrincipalTag/{tagParam.valueAsString}": True

principal = iam.AccountRootPrincipal().with_conditions({
    "StringEquals": string_equals

iam.Role(self, "MyRole", assumed_by=principal)

Explanation: since in this example we pass the tag name through a parameter, it can only be resolved during deployment. The resolved value can be represented in the template through a { "Ref": "TagName" }. However, since we want to use this value inside a aws:PrincipalTag/TAG-NAME IAM operator, we need it in the key of a StringEquals condition. JSON keys must be strings, so to circumvent this limitation, we use CfnJson to “delay” the rendition of this template section to deploy-time. This means that the value of StringEquals in the template will be { "Fn::GetAtt": [ "ConditionJson", "Value" ] }, and will only “expand” to the operator we synthesized during deployment.

Stack Resource Limit

When deploying to AWS CloudFormation, it needs to keep in check the amount of resources being added inside a Stack. Currently it’s possible to check the limits in the AWS CloudFormation quotas page.

It’s possible to synthesize the project with more Resources than the allowed (or even reduce the number of Resources).

Set the context key @aws-cdk/core:stackResourceLimit with the proper value, being 0 for disable the limit of resources.