AWS Compute Blog

Implementing Canary Deployments of AWS Lambda Functions with Alias Traffic Shifting

This post courtesy of Ryan Green, Software Development Engineer, AWS Serverless

The concepts of blue/green and canary deployments have been around for a while now and have been well-established as best-practices for reducing the risk of software deployments.

In a traditional, horizontally scaled application, copies of the application code are deployed to multiple nodes (instances, containers, on-premises servers, etc.), typically behind a load balancer. In these applications, deploying new versions of software to too many nodes at the same time can impact application availability as there may not be enough healthy nodes to service requests during the deployment. This aggressive approach to deployments also drastically increases the blast radius of software bugs introduced in the new version and does not typically give adequate time to safely assess the quality of the new version against production traffic.

In such applications, one commonly accepted solution to these problems is to slowly and incrementally roll out application software across the nodes in the fleet while simultaneously verifying application health (canary deployments). Another solution is to stand up an entirely different fleet and weight (or flip) traffic over to the new fleet after verification, ideally with some production traffic (blue/green). Some teams deploy to a single host (“one box environment”), where the new release can bake for some time before promotion to the rest of the fleet. Techniques like this enable the maintainers of complex systems to safely test in production while minimizing customer impact.

Enter Serverless

There is somewhat of an impedance mismatch when mapping these concepts to a serverless world. You can’t incrementally deploy your software across a fleet of servers when there are no servers!* In fact, even the term “deployment” takes on a different meaning with functions as a service (FaaS). In AWS Lambda, a “deployment” can be roughly modeled as a call to CreateFunction, UpdateFunctionCode, or UpdateAlias (I won’t get into the semantics of whether updating configuration counts as a deployment), all of which may affect the version of code that is invoked by clients.

The abstractions provided by Lambda remove the need for developers to be concerned about servers and Availability Zones, and this provides a powerful opportunity to greatly simplify the process of deploying software.
*Of course there are servers, but they are abstracted away from the developer.

Traffic shifting with Lambda aliases

Before the release of traffic shifting for Lambda aliases, deployments of a Lambda function could only be performed in a single “flip” by updating function code for version $LATEST, or by updating an alias to target a different function version. After the update propagates, typically within a few seconds, 100% of function invocations execute the new version. Implementing canary deployments with this model required the development of an additional routing layer, further adding development time, complexity, and invocation latency.
While rolling back a bad deployment of a Lambda function is a trivial operation and takes effect near instantaneously, deployments of new versions for critical functions can still be a potentially nerve-racking experience.

With the introduction of alias traffic shifting, it is now possible to trivially implement canary deployments of Lambda functions. By updating additional version weights on an alias, invocation traffic is routed to the new function versions based on the weight specified. Detailed CloudWatch metrics for the alias and version can be analyzed during the deployment, or other health checks performed, to ensure that the new version is healthy before proceeding.

Note: Sometimes the term “canary deployments” refers to the release of software to a subset of users. In the case of alias traffic shifting, the new version is released to some percentage of all users. It’s not possible to shard based on identity without adding an additional routing layer.


The simplest possible use of a canary deployment looks like the following:

# Update $LATEST version of function
aws lambda update-function-code --function-name myfunction ….

# Publish new version of function
aws lambda publish-version --function-name myfunction

# Point alias to new version, weighted at 5% (original version at 95% of traffic)
aws lambda update-alias --function-name myfunction --name myalias --routing-config '{"AdditionalVersionWeights" : {"2" : 0.05} }'

# Verify that the new version is healthy
# Set the primary version on the alias to the new version and reset the additional versions (100% weighted)
aws lambda update-alias --function-name myfunction --name myalias --function-version 2 --routing-config '{}'

This is begging to be automated! Here are a few options.

Simple deployment automation

This simple Python script runs as a Lambda function and deploys another function (how meta!) by incrementally increasing the weight of the new function version over a prescribed number of steps, while checking the health of the new version. If the health check fails, the alias is rolled back to its initial version. The health check is implemented as a simple check against the existence of Errors metrics in CloudWatch for the alias and new version.

GitHub aws-lambda-deploy repo


git clone
cd aws-lambda-deploy


# Rollout version 2 incrementally over 10 steps, with 120s between each step
aws lambda invoke --function-name SimpleDeployFunction --log-type Tail --payload \
  '{"function-name": "MyFunction",
  "alias-name": "MyAlias",
  "new-version": "2",
  "steps": 10,
  "interval" : 120,
  "type": "linear"
  }' output

Description of input parameters

  • function-name: The name of the Lambda function to deploy
  • alias-name: The name of the alias used to invoke the Lambda function
  • new-version: The version identifier for the new version to deploy
  • steps: The number of times the new version weight is increased
  • interval: The amount of time (in seconds) to wait between weight updates
  • type: The function to use to generate the weights. Supported values: “linear”

Because this runs as a Lambda function, it is subject to the maximum timeout of 5 minutes. This may be acceptable for many use cases, but to achieve a slower rollout of the new version, a different solution is required.

Step Functions workflow

This state machine performs essentially the same task as the simple deployment function, but it runs as an asynchronous workflow in AWS Step Functions. A nice property of Step Functions is that the maximum deployment timeout has now increased from 5 minutes to 1 year!

The step function incrementally updates the new version weight based on the steps parameter, waiting for some time based on the interval parameter, and performing health checks between updates. If the health check fails, the alias is rolled back to the original version and the workflow fails.

For example, to execute the workflow:

export STATE_MACHINE_ARN=`aws cloudformation describe-stack-resources --stack-name aws-lambda-deploy-stack --logical-resource-id DeployStateMachine --output text | cut  -d$'\t' -f3`

aws stepfunctions start-execution --state-machine-arn $STATE_MACHINE_ARN --input '{
  "function-name": "MyFunction",
  "alias-name": "MyAlias",
  "new-version": "2",
  "steps": 10,
  "interval": 120,
  "type": "linear"}'

Getting feedback on the deployment

Because the state machine runs asynchronously, retrieving feedback on the deployment requires polling for the execution status using DescribeExecution or implementing an asynchronous notification (using SNS or email, for example) from the Rollback or Finalize functions. A CloudWatch alarm could also be created to alarm based on the “ExecutionsFailed” metric for the state machine.

A note on health checks and observability

Weighted rollouts like this are considerably more successful if the code is being exercised and monitored continuously. In this example, it would help to have some automation continuously invoking the alias and reporting metrics on these invocations, such as client-side success rates and latencies.

The absence of Lambda Errors metrics used in these examples can be misleading if the function is not getting invoked. It’s also recommended to instrument your Lambda functions with custom metrics, in addition to Lambda’s built-in metrics, that can be used to monitor health during deployments.


These examples could be easily extended in various ways to support different use cases. For example:

  • Health check implementations: CloudWatch alarms, automatic invocations with payload assertions, querying external systems, etc.
  • Weight increase functions: Exponential, geometric progression, single canary step, etc.
  • Custom success/failure notifications: SNS, email, CI/CD systems, service discovery systems, etc.

Traffic shifting with SAM and CodeDeploy

Using the Lambda UpdateAlias operation with additional version weights provides a powerful primitive for you to implement custom traffic shifting solutions for Lambda functions.

For those not interested in building custom deployment solutions, AWS CodeDeploy provides an intuitive turn-key implementation of this functionality integrated directly into the Serverless Application Model. Traffic-shifted deployments can be declared in a SAM template, and CodeDeploy manages the function rollout as part of the CloudFormation stack update. CloudWatch alarms can also be configured to trigger a stack rollback if something goes wrong.


  Type: AWS::Serverless::Function
    FunctionName: MyFunction
    AutoPublishAlias: MyFunctionInvokeAlias
      Type: Linear10PercentEvery1Minute
        Fn::GetAtt: [ DeploymentRole, Arn ]
       - { Ref: MyFunctionErrorsAlarm }

For more information about using CodeDeploy with SAM, see Automating Updates to Serverless Apps.


It is often the simple features that provide the most value. As I demonstrated in this post, serverless architectures allow the complex deployment orchestration used in traditional applications to be replaced with a simple Lambda function or Step Functions workflow. By allowing invocation traffic to be easily weighted to multiple function versions, Lambda alias traffic shifting provides a simple but powerful feature that I hope empowers you to easily implement safe deployment workflows for your Lambda functions.