How [I think] AWS Lambda invocation works

July 16, 2019

While reading through Bref's source code and working with AWS Lambda, I found it to be important for me to get as much understanding of the process as I could. That usually helps me understand why an issue happens the way it does. For instance, deploying a php cli function that never calls Bref's lambda() helper leads to a never-ending execution. My understanding of the Lambda Function model explains why that happens. Here goes a lot of assumptions.

Timeline

Lambda

If we consider an API Lambda Function, it's usually executed by Api Gateway (APIGW) or Application Load Balancer. Regardless of which, the user will send a request using a service DNS so that the receiving end (APIGW or ALB) can then communicate with the Lambda Runtime API (an internal AWS service). I assume the Runtime service stores the payload of the request somewhere internally (DynamoDB?) and initiates the process to start a container. As soon as the container is up and running, it will have a chance to run any bootstrapping process and will end up entering a while(true) (Event Loop). Here, the code will make an HTTP call to the Lambda Runtime API asking if there's any work to be done. The answer seems obvious: of course there is work to be done, otherwise the container would not have started. This is where the actual Lambda Function starts it's execution and our code gets a chance to run. This also represents the end of the infamous Cold Start.

The Lambda code will finish it's execution and return a response. This response is sent back to the Lambda Runtime API via an HTTP Request that contains a unique identifier; this is how the Runtime knows to which request this response belongs to. In possession of a response, AWS's internal service now have the content that APIGW or ALB requested and can promptly reply to that, leading to the end of the Request / Response lifecycle.

Err... and the container?

Remember I said the container starts, bootstraps and enter an event loop? This means that the container sent a response back and started all over again. In the 2nd execution, the container makes another call to the Lambda Runtime to get some other work to do. If there's nothing more to do, the request will stay alive in a long-polling process waiting until there's an answer. I think two outcomes are possible: a new request comes in and the request finally gets a response so that it can start working on the next request or a new request never comes in. For the 1st option, it seems simple to imagine that the code will be executed again, send the response and start the loop all over. As soon as the container calls the Lambda Runtime to get a new payload, the billing time ends and the request stays locked in a long-polling. If it stays locked there for too long, AWS uses an external process to send a shutdown signal to the container, which means our code never gets executed again and the container is terminated.

Next time there's work to be done, a new container will have to face the bootstrapping process before it can work the payload.

What about non-api executions?

As I've said on another post, Lambda is extremely decoupled from it's executioner. If a Lambda function gets executed by API GW, ALB, SQS, SNS, S3, Aurora or direct invocation, it's all the Lambda Runtime's job to store the payload and start a container that will come back to the Runtime and ask for work to do. It's also the Lambda Runtime's job to know the context which a Lambda got executed. If an invocation comes from API GW or ALB, it should take the response and send it back as an HTTP Response. If it's source is SQS, it should delete the SQS Message. If it's SNS, I imagine there's nothing else to be done. If it's a direct invocation, it sends the response back to the client that invoked it.

Where does VPC cold start fits into this?

I'm not a network expert, but I can imagine the troublesome process of provisioning a secure channel between the Lambda container and a specific VPC on a specific account. This is never a generic network interface as each VPC is unique. Putting Lambda inside the VPC incur in the overhead of forcing AWS to provision such infrastructure. As announced on the last re:Invent 2018, the Lambda team is working on (as far as I understood) provisioning the network at Lambda creation time so that the network is always available and replacing this huge overhead of the cold start with a local NAT with a secure channel to the ENI provisioned on creation time, reducing the cost of starting a VPC-based Lambda.

Conclusion

This post contains a lot of assumptions. Some things make sense when you see how a provided runtime is programmed (such as Bref), but there are wild assumptions nonetheless. Even if they're wrong, they were still pretty useful insights for me to understand how my lambda can enter an endless execution if my code never makes the Lambda Runtime API call to get the reason why the container was started.

I also found it to be a bit weird that a service A (Lambda Runtime) would invoke a service B (start a container) and give it no context at all about it's purpose in life. But once we look underneath to see the reasoning for that, it makes it so beautiful. By letting the container come back to ask for something to do (breaking the Tell, Dont Ask pattern), AWS can leverage long-polling and an infinite while-loop to get the container to always come back to get more work to do from the Runtime.

I hope there are useful bits of information for anyone out there going through the challenges of learning Lambda Function, specially in the PHP context. If you have any feedback, find me on Twitter!

Cheers.