Machine learning

These examples live in the companion repository ubunye-ai-ecosystems/ubunye-examples. Run the commands from a clone of it (setup per place).

Training models and using them. Two examples: the full lifecycle done properly, and the same model built in three frameworks to prove the engine does not care which one you pick.

04 · The full lifecycle

Build features with a deterministic split, train, judge the model on rows it never saw, pass a quality gate, register, promote to production, then score new data with whatever model is in production. A model that fails the gate is never registered at all.

DatabricksLaptopDockerKubernetes

Open the notebooks under examples/04_ml_taxi_fare/notebooks/ in order, or run the whole chain:

databricks bundle run ml_taxi_fare --target dev

Seed once (see Start here), then run the three tasks in order:

TAXI_MODEL_STORE=/tmp/ubunye/model_store \
platforms/run_task.sh examples/04_ml_taxi_fare taxi_fare ml \
  feature_engineering model_training batch_inference

docker run --rm --entrypoint bash \
  -e TAXI_MODEL_STORE=/data/model_store ubunye-portable:ci -c \
  '. platforms/spark_env.sh && python platforms/seed.py && \
   platforms/run_task.sh examples/04_ml_taxi_fare taxi_fare ml \
     feature_engineering model_training batch_inference'

sed -e "s|{{NAME}}|ml|" \
    -e 's|{{ARGS}}|"examples/04_ml_taxi_fare", "taxi_fare", "ml", "feature_engineering", "model_training", "batch_inference"|' \
    platforms/k8s/job-template.yaml | kubectl apply -f -

10 · One model, three frameworks

The same model trained in scikit-learn, PyTorch and TensorFlow, behind three config files that are byte for byte identical. A build check keeps them identical. It ends in a leaderboard honest enough to say "too close to call" instead of crowning a half-percent lead.

DatabricksLaptopDockerKubernetes

databricks bundle run ml_frameworks --target dev

The three trainers run in parallel, then the judge waits for all three.

TensorFlow needs numpy<2, so install the pinned trio first:

pip install torch --extra-index-url https://download.pytorch.org/whl/cpu
pip install tensorflow "numpy<2" "protobuf<5"
BENCH_MODEL_STORE=/tmp/ubunye/model_store \
platforms/run_task.sh examples/10_ml_frameworks fare_bench ml \
  train_sklearn train_torch train_keras select_champion

The ml image already carries all three frameworks, pinned correctly:

docker run --rm --entrypoint bash \
  -e BENCH_MODEL_STORE=/data/model_store ubunye-ml:ci -c \
  '. platforms/spark_env.sh && python platforms/seed.py && \
   platforms/run_task.sh examples/04_ml_taxi_fare taxi_fare ml feature_engineering && \
   platforms/run_task.sh examples/10_ml_frameworks fare_bench ml \
     train_sklearn train_torch train_keras select_champion'

The examples repo has no ready manifest for this one, so here is the Docker command wrapped in a Job. It follows platforms/k8s/job-template.yaml, with the ubunye-ml:ci image and at least 5Gi of memory: three frameworks is a real appetite. Save it as job-ml-frameworks.yaml, then kubectl apply -f job-ml-frameworks.yaml:

apiVersion: batch/v1
kind: Job
metadata:
  name: ubunye-ml-frameworks
spec:
  backoffLimit: 0
  template:
    spec:
      restartPolicy: Never
      containers:
        - name: ubunye
          image: ubunye-ml:ci
          imagePullPolicy: Never # side-loaded into kind
          command: ["bash", "-c"]
          args:
            - >-
              . platforms/spark_env.sh && python platforms/seed.py &&
              platforms/run_task.sh examples/04_ml_taxi_fare taxi_fare ml feature_engineering &&
              platforms/run_task.sh examples/10_ml_frameworks fare_bench ml
              train_sklearn train_torch train_keras select_champion
          env:
            - name: BENCH_MODEL_STORE
              value: /data/model_store
            - name: DATA_DIR
              value: /data
          volumeMounts:
            - name: data
              mountPath: /data
          resources:
            requests: { memory: "5Gi", cpu: "2" }
            limits: { memory: "7Gi" }
      volumes:
        - name: data
          emptyDir: {}

Follow it with kubectl logs -f job/ubunye-ml-frameworks.