AWS Cdk

Troubleshooting: Deployment Failures

Table of Contents

• Overview

• Deploy Failure Root Cause Analysis

• Deadly Embrace (Cross-Stack Reference Deadlock)

• UPDATE_ROLLBACK_FAILED Recovery

• Non-Empty Bucket Deletion

Overview

This reference covers deployment-time failures — errors that occur after cdk synth succeeds and CloudFormation begins creating or updating resources. The CDK CLI error message is almost never the root cause; you MUST inspect CloudFormation stack events to find the actual failure.

Three error categories exist:

Deploy Failure Root Cause Analysis

The CDK CLI surfaces only a terse summary; the real cause is in the failed deployment, not the CLI output. You MUST work through these steps in order.

Step 1: Re-run with --verbose

cdk deploy $STACK --verbose

Prints every AWS API call, the change-set diff, and a fuller stack trace (-vv / -vvv for more).

Step 2: cdk diagnose (preferred, CDK CLI ≥ 2.1120.0)

cdk --unstable=diagnose diagnose $STACK

Inspects the failed deployment and prints the root cause with pointers back to the CDK source that caused it. It runs after the fact, so it also works for diagnosing CI/CD pipeline failures. Requires the --unstable=diagnose flag.

Step 3: CloudFormation events (fallback)

If cdk diagnose is unavailable (older CLI) or you need the raw stream:

aws cloudformation describe-events --stack-name $STACK --filters FailedEvents=true

describe-events groups events by operation ID and surfaces validation, provisioning, and hook-invocation errors — it supersedes describe-stack-events. The FIRST event in the output is the real root cause; later failures are rollback cascade.

Step 4: Read the ResourceStatusReason

Step 5: Service logs for Lambda / API Gateway / custom resources

CloudFormation only reports that a resource failed. The actual reason (e.g. a custom-resource Lambda threw) is in CloudWatch Logs:

• Lambda: /aws/lambda/<function-name>

• CodeBuild-in-pipeline: /aws/codebuild/<project>

• CloudFormation custom resources: the backing Lambda's log group.

EarlyValidationFailure specifically

Fails BEFORE the change set is submitted — a construct's validate() returned errors, a synth-time assertion tripped, or an addError annotation fired. The message names the exact property and constraint; fix it before redeploying.

If you have the awslabs aws-iac-mcp-server, its troubleshoot_cloudformation_deployment tool matches the failure event stream against 30+ known patterns and returns CloudTrail deep links — use it to shortcut Steps 2–4.

Deadly Embrace (Cross-Stack Reference Deadlock)

A deadly embrace occurs when Stack A exports a value that Stack B imports, and you then try to remove the export (or the resource behind it). CloudFormation refuses:

Export Stack1:ExportsOutputFnGetAtt-XXXX cannot be deleted as it is in use by Stack2

The deadlock is structural: a safe removal needs B deployed first (so it stops importing), but CDK orders A before B because of the dependency.

Every cross-stack reference has a strength:

• Strong (default) — uses Fn::ImportValue. CloudFormation blocks the producer from removing the export while any consumer still imports it.

• Weak — uses Fn::GetStackOutput. No coupling; the producer can be changed or deleted independently.

• Both — transitional state for migrating strong → weak.

Cross-account references are always weak (strong is unsupported cross-account).

Fix — reference strength (recommended)

CDK supports weakening the reference before removing the resource, with no manual exportValue hacks. You MUST do this as a three-deploy migration.

Weaken all references to a resource — CrossStackReferences.of(resource).produce():

import { CrossStackReferences, ReferenceStrength } from 'aws-cdk-lib';

// Deploy 1 — consumers move to Fn::GetStackOutput; the strong export stays
CrossStackReferences.of(bucket).produce(ReferenceStrength.BOTH);

// Deploy 2 — drop the strong export now that no consumer uses Fn::ImportValue
CrossStackReferences.of(bucket).produce(ReferenceStrength.WEAK);

// Deploy 3 — remove the resource or the reference entirely

Weaken a single reference — Stack.consumeReference():

import { Stack, ReferenceStrength } from 'aws-cdk-lib';

// Deploy 1 — wrap with consumeReference (defaults to BOTH)
new CfnOutput(consumer, 'BucketArn', { value: Stack.consumeReference(bucket.bucketArn) });

// Deploy 2 — switch to WEAK
new CfnOutput(consumer, 'BucketArn', {
  value: Stack.consumeReference(bucket.bucketArn, ReferenceStrength.WEAK),
});

// Deploy 3 — remove the resource or reference

(Use Stack.consumeListReference() for string-list references.)

Fix — legacy two-deploy (exportValue)

Use this only on CDK versions that lack ReferenceStrength. It MUST be done in exactly two deployments:

Deploy 1 — decouple the consumer, keep the export alive:

1. In consumer Stack B, remove the cross-stack reference (replace with a hardcoded value, SSM lookup, etc.).

2. In producer Stack A, add this.exportValue(resource.attribute) to keep the export alive during the transition.

3. Deploy both.

Deploy 2 — remove the export:

1. In Stack A, remove the this.exportValue() call (and the underlying resource if desired).

2. Deploy again.

You MUST NOT attempt to remove the export and the import in a single deployment.

Manual deploy ordering (cdk deploy -e)

If the consumer already stopped using the value and you control ordering yourself:

cdk deploy -e $CONSUMER_STACK   # deploy consumer first (drops the import)
cdk deploy -e $PRODUCER_STACK   # then producer, removing the export

-e / --exclusively deploys only the named stack and skips dependency reconciliation.

Prevention

• Default cross-stack references to weak for resources you expect to remove or replace. Set app-wide in cdk.json:

  Copy

  { "context": { "@aws-cdk/core:defaultCrossStackReferences": "weak" } }

• Keep stateful, long-lived resources in their own stack, separate from consumers.

• Use SSM Parameter Store as indirection (producer writes a parameter, consumer reads it) — no CFN export, no embrace.

UPDATE_ROLLBACK_FAILED Recovery

A stack enters UPDATE_ROLLBACK_FAILED when CloudFormation cannot roll back a failed update. The stack is wedged and MUST be recovered before any further operations.

First: was the failed deployment an express mode deployment?

If the failed operation was run with --express, the rollback-based recovery below does NOT apply. Express mode deployments cannot use the CloudFormation Rollback Stack API, and a standard CloudFormation deployment MUST NOT be used to recover a failed express mode deployment.

Recover by rolling forward instead — make another express mode deployment that resolves the failure, either by reverting the resource to its last successful state in code or by making another change that fixes the cause:

cdk deploy $STACK --express

Security note: A failed express deployment may have left resources partially configured — potentially without intended encryption, access controls, or policy attachments. Inspect the stack's resources (aws cloudformation list-stack-resources) before reusing the environment.

Audit note: Check CloudTrail logs for the failed deployment's API calls to determine which resources were created or modified and in what state they were left. This is the most reliable way to reconstruct exactly which create/update/delete operations succeeded or failed, since express mode reports the stack as failed without recording per-resource stabilization.

Running cdk deploy --express or cdk deploy --express --rollback against an already-failed stack applies the changes without first attempting a rollback, unlike cdk deploy without express mode. See fast-deployments.

Root causes

• Resource deleted out-of-band (e.g., manually deleted in the console).

• Insufficient IAM permissions for the rollback operation.

• Service quota exceeded.

• Resource operation timed out.

Recovery options

Option 1 — Standard rollback:

cdk rollback $STACK

Option 2 — Orphan stuck resources:

If a specific resource cannot be rolled back (e.g., it was deleted out-of-band), skip it:

cdk rollback $STACK --orphan $LOGICAL_ID

The resource is removed from the stack's state without attempting to delete or update it.

Option 3 — Force rollback:

cdk rollback $STACK --force

Post-recovery steps

After the stack returns to a stable state, you MUST:

1. Run cdk diff $STACK to understand the current drift.

2. Fix the root cause (restore deleted resources, fix IAM, request quota increase).

3. Redeploy: cdk deploy $STACK.

You SHOULD NOT leave a stack in a recovered-but-drifted state.

Non-Empty Bucket Deletion

Setting removalPolicy: cdk.RemovalPolicy.DESTROY alone MUST NOT be expected to delete an S3 bucket that contains objects. CloudFormation cannot empty a bucket during deletion. Versioned buckets are worse — delete markers and non-current object versions persist even after apparent object deletion, so the bucket can appear empty yet still fail to delete.

Fix

You MUST add autoDeleteObjects: true alongside the removal policy:

new s3.Bucket(this, 'MyBucket', {
  removalPolicy: cdk.RemovalPolicy.DESTROY,
  autoDeleteObjects: true,
});

autoDeleteObjects installs a custom resource Lambda that deletes all object versions and delete markers before CloudFormation attempts to delete the bucket.

You SHOULD only use this pattern in development or test stacks. Production buckets SHOULD retain the default removalPolicy: RETAIN.

Lambda Cannot Find Module at Runtime

These errors occur at Lambda invoke time, not during cdk synth. The function deploys successfully but fails when invoked.

Symptom

Cannot find module 'index'
Cannot find module 'aws-sdk'
Runtime.ImportModuleError: No module named 'requests'

Cause

• Wrong handler value (e.g., handler: 'handler' instead of handler: 'index.handler')

• aws-sdk v2 was removed from Node.js 18+ Lambda runtimes — code still imports it

• Python dependencies not bundled — Code.fromAsset() zips the directory without running pip install

Fix

• Fix handler to match your file and export: handler: 'index.handler'

• Migrate from AWS SDK v2 to v3: import { S3Client } from '@aws-sdk/client-s3'

• Remove externalModules: ['aws-sdk'] from bundling options if present

• For Python: use PythonFunction from @aws-cdk/aws-lambda-python-alpha which bundles pip dependencies automatically

API Gateway Multi-Stage

This is a construct design issue that manifests at deploy time, not a synth failure.

Symptom

Creating a RestApi produces only one stage. Adding extra Stage objects causes conflicts or duplicate deployments.

Cause

RestApi creates a Deployment and a default Stage automatically. Creating additional Stage objects without disabling the default causes conflicts.

Fix

Set deploy: false on the RestApi, then create Deployment and Stage objects explicitly:

const api = new apigateway.RestApi(this, 'Api', { deploy: false });
// ... define resources and methods ...
const deployment = new apigateway.Deployment(this, 'Deployment', { api });
new apigateway.Stage(this, 'Dev', { deployment, stageName: 'dev' });
new apigateway.Stage(this, 'Prod', { deployment, stageName: 'prod' });