A/B testing on AWS CloudFront with Lambda@Edge

Lorenzo Nicora
Feb 19, 2018 · 9 min read

AWS Lambda@Edge is still a young feature in AWS, Generally Available since July 2017.

Lambda@Edge allows running Lambda functions at Edge Locations of the CloudFront CDN. It means you may add “intelligence” in the CDN, without having to forward the request to a backend and losing benefits of content caching and geographical proximity with the client.

AWS Lambda@Edge and CloudFront CDN

Working at the new Buldit@Wipro Digital website (not yet published, so no link) we wanted the ability to run A/B testing.
The website is distributed by AWS CloudFront, so the solution was to use Lambda@Edge.

The Use Case

Imagine you have a static website or a Single Page Application served through the CDN. You want to experiment two versions with actual users. I’m focusing on SPA and static contents, as any non-cacheable, back-end generated content (e.g. non-cacheable Rest API requests) would not use the CDN regardless.

A/B testing doesn’t necessarily mean you have two equivalent versions taking 50% of the traffic each. I prefer calling the two versions “Main” and “Experiment”, rather than “A” and “B”, to make it clear we have a default and a secondary version. We’ll see why this is important.
Also, we are testing two complete versions of contents (our entire website or SPA). This is a more realistic use case rather than switching a single element, like a logo image.

A/B testing example (by Maxime Lorant from Wikipedia)

The fraction of traffic going to Experiment is just a parameter of the problem and the logic may be easily extended to multiple experiment versions… but let’s keep it simple.

In general, there are two approaches to A/B testing, client-side and server-side, depending on where you put the intelligence to switch traffic between the two versions.

I’m focusing on server-side A/B testing. This may be the only option when, for example, pages are served by an S3 Bucket with no brain. But is also a good choice if you don’t want to “pollute” neither your front-end nor your back-end with additional logic.

Let’s recap our scenario:

  • Front-end A/B testing: static content, SPA or any cacheable content in general.
  • Content is served through CloudFront CDN and we don’t want to lose benefits of the CDN: caching and geographical proximity.
  • We cannot (or do not want to) pollute our front-end with any switching logic. So the logic must be on the CDN not to lose it.
    Lambda@Edge is the way of putting logic on the CDN.
  • For simplicity, I’m serving content from S3 buckets directly. But the idea may be extended to any http/https source.
  • We have two complete versions of front-end: Main and Experiment. They are served from two separate S3 buckets (separate CloudFront Distribution Origins).
  • We want to randomly switch a fraction of users to Experiment.
  • We want a user to stay on the same version for the duration of the session (the browser “session”, remember we don’t have any login).

The last requirement is very important. We don’t want the user jumping from one version to another at every click, or worse, receiving parts of the same page from different versions.
There are a couple of examples of A/B Testing using in Lambda@Edge, in CloudFront Developer Guide and in the AWS blog. Both examples use a cookie to keep the user on a version, but both require an external logic to set the cookie. We said we don’t want to modify either front-end or back-end code.
So, these examples are not enough for us and the CDN must take care of setting the cookie.

How it works

Here is the working solution, using 3 functions.
We use a cookie, X-Source, to keep the user on a single version for the duration of the browser session. But the client might not have it when it lands on the website…

Stable A/B testing with Lambda@Edge
  1. The browser request is directed to the closest AWS Edge Location. The request may contain the X-Source cookie.
  2. The Viewer Request Lambda@Edge function gets triggered on every request. If the cookie is not in the request, it rolls dice to decide which version to send the user to and adds the cookie accordingly.
  3. The Distribution decides whether the request is a cache hit. We are forwarding the X-Source cookie and the cookie is part of the cache key (see CloudFront settings, below).
  4. If the request is a cache-miss, it triggers the Origin Request function. This function is triggered only on cache-miss and introduces no overhead on hits. By default, the Distribution forwards cache-misses to the Main Origin. If the cookie points to Experiment, the request Origin is modified.
  5. The content is served by either S3 bucket. S3 completely ignores the cookie.
  6. The response from the Origin triggers the Origin Request function. This happens only on cache misses. It adds a Set-Cookie header to set X-Source. Remember the cookie might have been added by the Viewer Request function and not by the browser.
  7. The decorated response, including the Set-Cookie header, is cached by the Distribution. The cache key is the object URI and the X-Cookie.
  8. The response is returned to the browser from the Edge location.
  9. The browser complies with the Set-Cookie headers, setting the cookie. The version will be stable for the duration of the session.
  10. In case of a cache hit, the response would have been returned immediately from the Edge Location and it would include the Set-Cookie header also cached (7).

This solution allows leveraging CDN caching, keeping both versions of content in the cache.

The additional overhead of the Viewer Request function, executed on every request, is negligible. We are using a single function for all contents, so it is “hot” most of the time, avoiding “cold-start” latency (though Node.js cold-start is fast and Lambda@Edge cold-start is even faster).
I observed a function execution time <1 ms most of the times, with few outliers of 15..20ms.

We don’t need to make any change to the application.

The default behaviour is serving the content from Main. This means that turning the A/B on and off just means attaching and detaching a single function: Viewer Request. The switching is (almost) atomic for a user, (usually) hitting the same Edge Location all the times, regardless the distributed and very eventually consistent nature of the CDN.

CloudFront Distribution settings

The CloudFront Distribution has two Origins, pointing the Main and Experiment sources. Both S3 bucket in our example and both with access restricted by Origin Access Identity.

CloudFront Distribution, Origins

The Default(*) Behaviour points the Main Origin.
It must Forward the X-Source cookie, as a whitelist. This cookie becomes part of the cache key. You don’t need a Behaviour for the Experiment Origin, as the Origin is switched dynamically by a function.

CloudFront Distribution, Default Behaviour settings

Lambda functions

I’m using Node.js for functions.

This function intercepts every client request, before deciding whether it is a cache hit.

The function looks for the X-Source cookie. If missing, it rolls dice to decide the version to serve and adds the cookie to the request.

This function intercepts requests on cache misses only.

If X-Source is present and points to Experiment, the Origin is changed. Otherwise it remains untouched, pointing to Main, as by Behaviour settings.

The Host header must be changed accordingly, or the execution would fail, complaining about “The request signature we calculated does not match the signature you provided”.

This function intercepts responses on cache misses only.

The event passed to the functions contains both the request decorated by Viewer Request and the response from the Origin.

We add a Set-Cookie header to be sure the browser will send the same cookie on every future request. The cookie Path guarantees three is a single cookie, regardless the path of the object.

You probably noticed all parameters are hardwired in the code. This is not due to my bad Node.js code. Lambda@Edge does not support passing configuration to functions as environment variables. Not to hardwire parameters, you must inject them at function build and deploy time.

Lambda@Edge functions require an execution role the same way as “normal” Lambda. If you are not accessing any external resources (you’d better not to) the standard AWSLambdaBasicExecutionRole will suffice. It allows functions to send logs to CloudWatch.

Association between Lambda and CloudFront Distribution is at Behaviour level.

Modify the Behaviour and wait for the Distribution change to completely propagate before trying any test.

All functions must be published with a numbered version ($LATEST is not supported). This is a documented, but not so obvious limitation.


Gotcha! (by Bill Larkins, from Wikipedia)

I reckon Lambda@Edge are quite new. Official documentation is sparse at best, and examples are usually naive and not much realistic. In a head-breaking cut&try process I gathered a number of gotchas. Some of them are documented, well hidden or not obvious at all (at least to me).

Lambda functions must be created in us-east-1 Region to be attached to CloudFront Distributions.

This is the same restriction that applies to other attached to the global CDN, like SSL certificates stored in Amazon Certificate Managers to use https on custom domains.

As already mentioned, the configuration must be hardwired in the code, as no environment variable is supported.

Lamba@Edge also has some other noteworthy limitations.

Execution logs from Lambda@Edge functions go to CloudWatch, as for normal Lambda functions, but the Region logs end up in is not obvious.

Logs go to the Region of the Edge Location the function is executed in.

This is the Edge Location the client request is directed to. It is usually the Location closest to the client, but not necessarily. I saw my requests from London switching from LHR (West London) Edge Location to AMS (Amsterdam). As a result, Lambda@Edge logs switching between eu-west-2 and eu-central-1 Regions.

Log Groups get qualified with the name of Region the function is stored in (not where it is executed). So the Log Group is always/aws/lambda/us-east-1.<function-name>…except when function executes in us-east-1, as it happens when you run it from AWS Console for debugging. In this case the Log Group /aws/lambda/<function-name>

If you change the Origin, don’t forget to change the `Host` header accordingly, to avoid “The request signature we calculated does not match the signature you provided” errors.

The request Origin may be changed only in Origin Request, not in Viewer Request. This way the response get always cached.

Do not forget to set up the Default Behaviour to forward the element you are using for deciding the Origin (the cookie, in our case), or the Origin Request function would not receive it.

This also brings to the next gotcha…

A forwarded cookie becomes part of the cache key, along with the object URI, regardless the Origin is ignoring it (e.g. S3).

Invalidation is by URI only. There is no way of invalidating a single version of an object (e.g. associated with one X-Source=Experiment)

This is probably the most annoying gotcha.

You obviously cannot delete a Lambda function while associated with a CloudFront Distribution. What is not obvious is you have to wait the function is completely removed from CDN replication before deleting it.

Until very recently there was no way of removing replicated functions, making users very… disappointed.

At the time of writing (February 2018), when you remove all associations to CloudFront Distributions, function replica are dropped after… a while.

I cannot exactly quantify the time it takes. It is not documented and there is no way of verifying the replication status through AWS Console or API/CLI. It does not look related to the Distribution deployment and I was able to delete detached functions after a variable time between 30 minutes to 24 hours.

This makes Lambda used at Edge not easily manageable by stateful infrastructure provisioning tools, like CloudFront or Terraform.


Lambda@Edge opens a number of very interesting use cases. Unfortunately, documentation is still poor, examples not so useful and tooling support non-existent or sparse.

The many limitations reduce possible uses. Also, you should avoid any heavyweight processing or access to external resources when running from Edge, or you give up any benefit of using a CDN.

Regardless limitations, implementing A/B testing is not so complex.


we enable + accelerate engineering transformation


We are a global engineering & co-innovation consultancy specializing in solving highly complex business problems. We leverage engineering best practices, tools, methodologies and new ways of working to deliver true business outcomes.

Lorenzo Nicora

Written by


We are a global engineering & co-innovation consultancy specializing in solving highly complex business problems. We leverage engineering best practices, tools, methodologies and new ways of working to deliver true business outcomes.

Medium is an open platform where 170 million readers come to find insightful and dynamic thinking. Here, expert and undiscovered voices alike dive into the heart of any topic and bring new ideas to the surface. Learn more

Follow the writers, publications, and topics that matter to you, and you’ll see them on your homepage and in your inbox. Explore

If you have a story to tell, knowledge to share, or a perspective to offer — welcome home. It’s easy and free to post your thinking on any topic. Write on Medium

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store