jcpsantiago/hazal

Test custom Sagemaker containers from the comfort of your laptop

hazal🧙‍♀️

AWS Sagemaker relies on docker containers for its pipelines. This means it's possible to create custom containers using e.g. other languages than Python or algorithms not available off-the-shelf in AWS. Even though it's easy to test each container in isolation, it is harder to test the whole pipeline in one go. hazal helps by orchestrating the communication between containers.

The usual disclaimers of beta-quality software apply.

Usage

You need to install janet on your machine first.

hazal expects a configuration file with the following structure, where each struct is a container. Order is important:

[{:type "single"               # type of container: single or multi-model
  :container "sagemaker-pre"}  # name of the container (docker run --name ...)
{:type "multi"
 :container "sagemaker-inf"}]

Such a file could live in the main repository of a project e.g. where models are trained locally. I call mine containers.jdn, but it doesn't really matter. Then, to use hazal and test the pipeline:

• Launch the docker containers (hazal won't do this for now)
• Launch hazal with janet main.janet <path to config> (or build the binary with jpm build)
• POST the payload expected by the first sagemaker container to localhost:9001/pipeline, or wherever hazal is running
• If you are using a multimodel container in the pipeline, specify which model to use for the prediction with a query parameter localhost:9001/pipeline?model=allseeingeye

TODO

• Define configuration structure
• Chain an arbitrary number of containers
• Differentiate between single and multi-model containers
• Send request to load model, if necessary
• Inform if the model does not exist
• Get host and port from docker ps
• If a container fails, pass the response and inform which step failed

See also

• MultiModel server in R
• MultiModel server in Clojure