The Code* family builds a model pipeline

Lesson 1 of 4 in MLOps Pipelines: CI/CD for Models.

The premise of MLOps automation: the next model version should ship itself. AWS's CI/CD parts bin is the Code* family, and the exam wants each member's one-line job description. CodePipeline orchestrates — stages and actions from commit to production. CodeConnections wires external git (GitHub, GitLab, Bitbucket) in as the source. CodeBuild runs the work: tests, container builds, kicking off training. CodeDeploy automates rollouts with traffic-shifting strategies. CodeCommit — AWS's own git hosting — is still named in scope, with an honest asterisk below.

A model pipeline, stage by stage

  1. Source

    A push to the main branch arrives via CodeConnections (or CodeCommit).

  2. Build & test (CodeBuild)

    Unit tests, data validation, container build to ECR — buildspec.yml defines it.

  3. Train & evaluate

    The pipeline triggers training (a SageMaker pipeline or training job) and computes evaluation metrics.

  4. Quality gate

    Does the candidate beat the baseline on held-out data? Fail → stop, nothing ships.

  5. Register version

    The passing model lands in the Model Registry, pending approval.

  6. Deploy

    CodeDeploy / a CloudFormation action / an update-endpoint step rolls it out — strategies next lesson.

What makes ML CI/CD more than software CI/CD: the tests
Test layerWhat it checksWhen it blocks

Data tests

Schema, ranges, nulls, distribution of the incoming training data

Before training — bad data never reaches the trainer

Model quality gate

Candidate metrics vs the current baseline on held-out data

Before registration — a worse model never becomes a version

Integration tests

A staging endpoint answers correctly, within latency, end to end

Before production traffic shifts

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.