Ansible for AI-Native Development — CI/CD Pipelines for ML

Introduction

AI-native software development — where coding, testing, and app building are increasingly AI-led — tops Gartner's 2026 trends. This shifts infrastructure needs: teams need ML CI/CD pipelines, model registries, experiment tracking, feature stores, and LLM evaluation frameworks. Ansible automates the entire MLOps platform stack.

MLOps Platform Stack

Component	Tool	Purpose
Experiment Tracking	MLflow, W&B	Track runs, metrics, artifacts
Model Registry	MLflow, DVC	Version and stage models
Feature Store	Feast, Hopsworks	Serve ML features
Pipeline Orchestration	Airflow, Prefect	Schedule training jobs
Model Serving	Triton, BentoML	Serve predictions
Monitoring	Evidently, Whylogs	Detect data/model drift

Deploy MLflow (Experiment Tracking + Registry)

---
- name: Deploy MLflow platform
  hosts: mlops_servers
  become: true
  vars:
    mlflow_port: 5000
    artifact_store: /opt/mlflow/artifacts
    db_uri: "postgresql://mlflow:{{ vault_db_password }}@localhost:5432/mlflow"

  tasks:
    - name: Install MLflow
      ansible.builtin.pip:
        name:
          - mlflow[extras]
          - psycopg2-binary
          - boto3
        virtualenv: /opt/mlflow/venv

    - name: Create artifact store
      ansible.builtin.file:
        path: "{{ artifact_store }}"
        state: directory
        owner: mlflow
        mode: "0755"

    - name: Create MLflow database
      community.postgresql.postgresql_db:
        name: mlflow
        state: present

    - name: Create MLflow database user
      community.postgresql.postgresql_user:
        name: mlflow
        password: "{{ vault_db_password }}"
        db: mlflow
        priv: ALL
      no_log: true

    - name: Deploy MLflow systemd service
      ansible.builtin.copy:
        content: |
          [Unit]
          Description=MLflow Tracking Server
          After=network.target postgresql.service

          [Service]
          Type=simple
          User=mlflow
          ExecStart=/opt/mlflow/venv/bin/mlflow server \
            --backend-store-uri {{ db_uri }} \
            --default-artifact-root {{ artifact_store }} \
            --host 0.0.0.0 \
            --port {{ mlflow_port }}
          Restart=always
          Environment=MLFLOW_TRACKING_URI=http://0.0.0.0:{{ mlflow_port }}

          [Install]
          WantedBy=multi-user.target
        dest: /etc/systemd/system/mlflow.service
      notify:
        - Reload systemd
        - Restart MLflow

    - name: Start MLflow
      ansible.builtin.service:
        name: mlflow
        state: started
        enabled: true

    - name: Verify MLflow is running
      ansible.builtin.uri:
        url: "http://localhost:{{ mlflow_port }}/api/2.0/mlflow/experiments/search"
        method: POST
        body_format: json
        body: {}
      register: mlflow_check

  handlers:
    - name: Reload systemd
      ansible.builtin.systemd:
        daemon_reload: true
    - name: Restart MLflow
      ansible.builtin.service:
        name: mlflow
        state: restarted

Deploy Feature Store (Feast)

- name: Deploy Feast feature store
  hosts: feature_store
  become: true
  tasks:
    - name: Install Feast
      ansible.builtin.pip:
        name:
          - feast[redis,postgres]
        virtualenv: /opt/feast/venv

    - name: Deploy feature repository
      ansible.builtin.git:
        repo: "{{ feature_repo_url }}"
        dest: /opt/feast/repo
        version: main

    - name: Deploy Redis for online store
      community.docker.docker_container:
        name: feast-redis
        image: redis:7-alpine
        state: started
        restart_policy: unless-stopped
        ports: ["6379:6379"]

    - name: Configure feature store
      ansible.builtin.template:
        src: feature_store.yaml.j2
        dest: /opt/feast/repo/feature_store.yaml

    - name: Apply feature definitions
      ansible.builtin.command:
        cmd: /opt/feast/venv/bin/feast apply
        chdir: /opt/feast/repo
      register: feast_apply
      changed_when: "'Created' in feast_apply.stdout or 'Updated' in feast_apply.stdout"

    - name: Deploy feature serving service
      ansible.builtin.copy:
        content: |
          [Unit]
          Description=Feast Feature Server
          After=network.target redis.service

          [Service]
          Type=simple
          User=feast
          WorkingDirectory=/opt/feast/repo
          ExecStart=/opt/feast/venv/bin/feast serve --host 0.0.0.0 --port 6566
          Restart=always

          [Install]
          WantedBy=multi-user.target
        dest: /etc/systemd/system/feast-server.service
      notify: Restart Feast

ML CI/CD Pipeline

- name: Deploy ML CI/CD infrastructure
  hosts: cicd_servers
  become: true
  tasks:
    - name: Deploy Airflow for pipeline orchestration
      community.docker.docker_compose_v2:
        project_src: /opt/airflow
        state: present
      vars:
        compose_file: |
          services:
            webserver:
              image: apache/airflow:2.9.0
              ports: ["8080:8080"]
              volumes:
                - /opt/airflow/dags:/opt/airflow/dags
              environment:
                AIRFLOW__CORE__EXECUTOR: LocalExecutor
                AIRFLOW__DATABASE__SQL_ALCHEMY_CONN: "{{ vault_airflow_db }}"

    - name: Deploy ML pipeline DAGs
      ansible.builtin.template:
        src: "dags/{{ item }}.py.j2"
        dest: "/opt/airflow/dags/{{ item }}.py"
      loop:
        - training_pipeline
        - evaluation_pipeline
        - deployment_pipeline
        - data_validation

    - name: Deploy model evaluation framework
      ansible.builtin.pip:
        name:
          - deepeval
          - ragas
          - promptfoo
        virtualenv: /opt/ml-eval/venv

    - name: Deploy LLM evaluation config
      ansible.builtin.template:
        src: eval-config.yml.j2
        dest: /opt/ml-eval/config.yml

    - name: Schedule automated model evaluation
      ansible.builtin.cron:
        name: "Daily model evaluation"
        hour: "6"
        minute: "0"
        job: >
          cd /opt/ml-eval &&
          venv/bin/python run_eval.py --config config.yml
          >> /var/log/ml-eval.log 2>&1

Model Drift Monitoring

- name: Deploy model drift monitoring
  hosts: monitoring
  become: true
  tasks:
    - name: Deploy Evidently for drift detection
      community.docker.docker_container:
        name: evidently
        image: evidentlyai/evidently-service:latest
        state: started
        ports: ["8085:8080"]
        volumes:
          - /opt/evidently/config:/config

    - name: Configure drift monitoring
      ansible.builtin.template:
        src: drift-config.yml.j2
        dest: /opt/evidently/config/monitoring.yml

    - name: Deploy drift alert rules
      ansible.builtin.template:
        src: drift-alerts.yml.j2
        dest: /etc/prometheus/rules/drift-alerts.yml
      notify: Reload Prometheus

{# templates/drift-alerts.yml.j2 #}
groups:
  - name: model_drift
    rules:
      - alert: DataDriftDetected
        expr: evidently_data_drift_score > 0.3
        for: 1h
        labels: { severity: warning }
        annotations:
          summary: "Data drift detected in {{ $labels.model }} (score: {{ $value }})"

      - alert: PredictionDriftDetected
        expr: evidently_prediction_drift_score > 0.2
        for: 30m
        labels: { severity: critical }
        annotations:
          summary: "Prediction drift in {{ $labels.model }} — retrain needed"

Complete MLOps Playbook

---
# site.yml — Full MLOps platform
- import_playbook: playbooks/mlflow.yml        # Experiment tracking
- import_playbook: playbooks/feature-store.yml  # Feature serving
- import_playbook: playbooks/pipelines.yml      # CI/CD orchestration
- import_playbook: playbooks/model-serving.yml  # Inference
- import_playbook: playbooks/monitoring.yml     # Drift detection

# Deploy full MLOps stack
ansible-playbook site.yml -e @vars/production.yml

# Update just pipelines
ansible-playbook playbooks/pipelines.yml --tags dags

# Deploy new model version
ansible-playbook playbooks/model-serving.yml -e "model_version=v2.3.0"

Conclusion

AI-native development in 2026 requires MLOps infrastructure that's as mature as traditional DevOps. Ansible automates the entire platform: MLflow for experiment tracking and model registry, Feast for feature stores, Airflow for pipeline orchestration, Evidently for drift monitoring, and evaluation frameworks for LLM quality. The pattern: each MLOps component is a deployable service managed by Ansible roles, composed into a full platform with a single site.yml. As AI becomes the default way to build software, this infrastructure automation is table stakes.