Deprecation Notice

I recommend now using Jot.jl, which does roughly the same thing as this but operates as a Julia package, has more functionality, and is better documented.

Jot is for creating AWS Lambda images in Julia

AWS does not provide native support for Julia, so functions must be put into containers which implement AWS's Lambda API, and uploaded to AWS ECR. This repo aims to reduce this to a simple, customizable and transparent process, which results in a low-latency Lambda function:

1. Enter your configuration for your AWS account and the function to be created in the config.json file
2. Add your function code to the react_to_invocation function in function/function.jl
3. Build the image locally: ./jot buildimage. This will build your docker container. It also adds two directories: image - containing the Dockerfile, and the files that will be added to the Docker image - and scripts, containing bash scripts that will be usful for managing the image. Amongst these are scripts to:
4. (Optionally) test the container locally by invoking it with a sample JSON, eg bash scripts/test_image_locally.sh '{"a":10, "b":15, "operation":"+"}'
5. Define the function on AWS Lambda - bash scripts/push_image_to_ecr_and_create_lambda_function.sh to push the docker image to Amazon ECR and create the associated function in AWS Lambda.

config.json

This is the config file for the Lambda image. It is divided into three parts (aws, image, and lambda_function, for your aws details, the docker image definition, and the AWS Lambda function respectively. Most of these fields have sensible defaults and can be left as they are. The fields are as follows (* denotes that a field should be changed from its default):

aws

• * account_id is your 12-digit AWS ID. See below for how to find this.
• * region is your AWS region code. See below for how to find this.
• role is the name for the AWS role that will execute the lambda function. It has no functional effect. If this role does not already exist, it will be created.

image

• name is your chosen name for your Julia image. It will have the AWS Account ID, the region and other data prepended to it when built. The name has no functional effect.
• tag is the tag for your image. If you are not using version numbers, just use latest.
• base is the name of the Julia base image, by default 1.6.0. A full list can be found at docker hub
• * dependencies is a list of Julia packages to add to the image. This should just be a list of strings that will be used with Pkg.add() - for example, ["DataFrames"] for DataFrames.jl, or ["DataFrames", "Distributions"]. These will be added to the image during the image build and precompiled.
• runtime_path is the path that will be created on the docker image to store the files used at runtime, including the julia file containing your function.
• julia_depot_path is the path that will be created on the docker image to act as the Julia depot path.
• julia_cpu_target is the JULIA_CPU_TARGET variable that will be used when compiling the docker image. If you change this, note that if it is incompatible with the hardware being used by AWS Lambda, the function may not work.

lambda_function

• * name is the name for your lambda function. The function will be re-built each time the push_image_to_ecr_and_create_lambda_function.sh script is run. Note that this begins by erasing any existing functions in AWS Lambda with the same name.
• timeout is the period that Lambda will wait for the container to respond before timing out. Lambda's default value (3 seconds) is too small for a cold-started Julia container that has not been pre-packaged. 20 seconds should be sufficient.
• memory_size is the allocated memory to run the container with. Since Lambda charges per GB-second, you want this to relatively low. If it is set too low, Julia will throw a Segmentation Fault.

Adding Function Code

The AWS Lambda API permits the container to return two values from a function call - a response, or an error. These are both JSONs, of no defined format.

The only place you will need to add your function code is in the react_to_invocation function, in function/function.jl. This function can return either an InvocationResponse (a return value), or an InvocationError (an error value). The InvocationResponse struct has a single attribute, response::String, to which the responding function should pass a JSON-compatible string. The remaining code (in template/image/_runtime/runtime.jl) will handle the return value appropriately, depending on whether it is a response or an error. The Invocation object contains the full details for the invocation, but it is likely that you will only be interested in the Invocation.body attribute, which contains a JSON of the original function argument.

Note that AWS does not specify the format of its errors, and the fields of InvocationError that has been pre-defined in function/function.jl are arbitrary and may be changed without issue.

Testing the container locally

AWS provides an RIE (Runtime Interface Emulator) to allow you to test Lambda images locally before pushing them to AWS. This runs the container in a proxy of the environment that will be used on AWS servers. By running the run_image_locally.sh script, the container will be started locally (on port 9000). It will run non-detached, so to test it, open up another terminal window, then run the test_image_locally.sh script, passing it a JSON string.

If you have not defined your own lambda function, the default one does simple addition or subtraction based on passed JSON, so try bash scripts/test_image_locally.sh '{"a":10, "b":15, "operation":"+"}', which should return 25.

Defining the AWS Lambda function

When ready, run bash scripts/push_image_and_create_function.sh to run a script that will push the image to AWS Elastic Container Registry, and then create a Lambda function based on that image. You can log into AWS and go to the Lambda console to observe the function and test it.

Building the files, or building the files and the image

The ./jot executable comes with multiple commands, amongst which are buildfilesonly and buildimage. These two offer different routes for building the docker image:

• buildimage builds the docker image in full for you
• buildfilesonly adds the files to be added to the image, and a Dockerfile, to the image folder, but does not build the image. This allows you to edit any of these files prior to building. When ready to build, run bash ./image/build_image.sh to invoke the image/build_image.sh script, which will build the image locally. If you edit any files in the image directory, note that running either ./jot buildimage or ./jot buildfilesonly will overwrite your changes from the template folder. If you would like to make persistent changes to any of these files, you should edit the template folder. After going through either of these processes, you should see the image in the output of docker image ls. The image name will contain your AWS Account number as well as the image name and tag defined in config.json.

Within the Dockerfile, the build_runtime.jl Julia script will be called. Amongst other things, this will use Pkg to add the runtime from a separate repo - harris-chris/julia-lambda-runtime.

Another important option is the --packaged command-line argument. An image built with --packaged uses PackageCompiler.jl to create a custom system image for the Lambda runtime and for your function. This will increase image build times, but decrease Lambda function run-times. It is therefore recommended for production, but not for testing. More details on how to optimize an image here.

The image directory

The ./image directory contains two things:

• the Dockerfile for the image, and a build_image.sh script to create the image using the Dockerfile. The Dockerfile is created entirely programmatically and does not exist in the template folder.
• a representation of the directory structure that will be added to the docker image. This includes, at minimum, the directory for the function runtime files, and the directory for the julia depot files. By default, these will be image/var/runtime and image/var/julia respectively. This representation is taken from the template/image directory.

The scripts directory

The ./scripts folder is created the first time you run a build, and contains bash scripts that are useful for working with the build. Their names are fairly self-explanatory.

Where do the scripts and image directories come from?

The template folder contains templates for both the scripts and image folder that are generated in the project root folder. Each time the ./jot executable is invoked with a build... command like buildimage or buildfilesonly, the following process takes place:

• the existing scripts and image in the root folder are deleted.
• they are then replaced by their versions in the template folder.
• all script files in the template folder (identified by having a shebang) have their placeholders for configuration variables replaced - for example, the run_image_locally.sh script looks like this in the template/scripts folder:

docker run \
  -p 9000:8080 \
  -it $(image.image_uri_string)

After it is copied to ./scripts, $(image.image_uri_string) is replaced with the image uri, so that it becomes, say,

docker run \
  -p 9000:8080 \
  -it 111111111111.dkr.ecr.ap-northeast-1.amazonaws.com/julia-lambda:latest

• any special folder names in the newly-created image directory, like _runtime or _depot, are changed to the their appropriate values from the config.json.
• finally, the lambda function itself is copied from ./function to the runtime folder - which is defined in config.json but is by default ./image/var/runtime.

If you would like to make persistent changes to the scripts, or the files that are ultimately included in the docker image, you should make them in the template folder, in which case they will be included in future builds created from this folder.

How to reduce function run-times

When your AWS Lambda function is invoked, your container may be active or (if it has not been used for several minutes) inactive. If inactive, this is referred to as a cold start. A cold start is undesirable for all Lambda run-times, but perhaps moreso for Julia because there is a short delay while Julia initiates. To mitigate this, the --packaged option is provided. If this is used, then PackageCompiler.jl will be used to build a system image - this is a pre-constructed definition of what the Julia session containing your runtime will look like. To build this system image, it is necessary to define an 'example run' of the function. PackageCompiler uses this example run to see which methods are used during the run and so can be compiled in advance. An existing example run is provided in the ./function/precompile_runtime.jl file. This example can be enhanced by, say, adding a more realistic invocation, which may improve performance further.

AWS-Craft

Finding your user details

Your account ID and your region can be found on the AWS navigation bar at the top of the screen: After logging in to AWS, your user ID can be found by clicking on your username: And your region code is listed next to the region name:

Troubleshooting

• Getting a Segmentation Fault while running in AWS: if the allocated memory for the Lambda function (defined in the MemorySize field of the config.json) is too low, Julia will not fail gracefully and will throw a Segmentation Fault. The default Lambda memory size (100mb) will cause this to occur.

Any problems, please log them as an issue on github!