Skip to content

Bonus Assignment: Deploy and automate one small model

This optional, ungraded assignment reviews Domain 3 by taking one small model from a hosting decision through to an automated release.

It is hands-on: the steps below are meant to be run in your own AWS account or a sandbox account.

Disclaimer

This project is for educational purposes only and might lead to AWS service charges. You assume any risk associated with this project. You are responsible for any costs associated with your AWS account.

The scenario

AnyCompany Logistics has a trained model that estimates delivery delay, and a separate internal assistant that answers dispatcher questions from the company's operating procedures. Both are finished and neither is in production. Dispatchers need the delay estimate while they are on a call, and the overnight planning system needs the same estimate applied to the whole next-day route file. You are deciding how each one is served, provisioning it, and automating the release.

Steps

Task 3.1: Manage deployment infrastructure

  • Write down, for each of the three workloads, who is waiting for the answer, how fast they need it, and how the work arrives. Name the inference option each set of answers points at before you open a console.
  • Deploy the delay model to a real-time endpoint. Note every field the endpoint configuration made you decide, and which of those fields would disappear on a serverless endpoint.
  • Say what would have to change to serve the same model as a batch job over the route file, and name what you would keep from the real-time deployment and what you would discard.
  • For the assistant, decide whether it needs a knowledge base, an agent, or both, and write one sentence saying what each one would be responsible for.

Task 3.2: Provision and configure resources

  • Register the endpoint as a scalable target and attach a policy. Choose the metric deliberately, and write one sentence on why the metric you picked reacts on the right timescale for dispatcher traffic.
  • Place the endpoint in private subnets and then find out what breaks. Add the private path it needs to reach its artifacts, and record what you had to add and why the endpoint could not start without it.
  • Create a small knowledge base over a handful of documents. Note the two settings you could not change afterward without reindexing, and say why they are locked.
  • Run one query with the default retrieval configuration, then change exactly one lever, either the result count, the search type, or a metadata filter. Record which complaint that lever would fix and which it would not.

Task 3.3: Implement automated orchestration and CI/CD

  • Register the trained model in a registry as a version inside a model group, and set its approval status. Note what a deployment step can now read that it could not read before.
  • Update the endpoint with a canary traffic shift, and configure an alarm for the rollout to watch. Then remove the alarm and say what the rollout would do differently.
  • Sketch the trigger that would retrain this model, choose between a schedule and an event, and justify the choice in one sentence against how the delivery data actually arrives.
  • Store the assistant's prompt as a versioned resource with a variable in it, create a second variant, and write down how you would decide between them in a pipeline rather than by reading both.

Check yourself

Reading the requirement

Could you defend each inference option choice to someone who proposed putting all three workloads on one real-time endpoint?

Network before permissions

Could you explain why a private endpoint failed to start, in terms of a missing path rather than a missing permission?

What the alarm is for

Could you say in one sentence what a canary rollout does when no alarm is configured, and why that is worse than no rollout strategy at all?

Versioning the whole change

Could you name every artifact in this scenario that needs a version, and say what an auditor could not reconstruct if one of them were missing?

Find your gap

If any step felt unclear, return to the skills in that task's section before moving on. Note the gap so you know what to review.

If you have no AWS account

Every step has a paper equivalent. Write out the endpoint configuration fields instead of creating the endpoint, draw the subnets and the private paths instead of provisioning them, list the two locked knowledge base settings instead of creating an index, and describe the canary shift and its alarm instead of running an update. The decisions are the review; the console is only where they become concrete.

Clean up

Delete anything you created, so nothing is left running or stored: delete the endpoint and its endpoint configuration, remove the scaling policy and any alarm, delete the knowledge base and the vector index behind it, empty and remove any bucket you created, and delete the private endpoints and the subnets if you created them for this exercise. If you used a generative AI tool to produce sample documents, clear its conversation history so the sample data is not retained.