Tutorial 6: Deploy & Run Inference

This tutorial deploys the inference task graph and runs batch predictions using the trained model.

What you will learn

• How the inference DAG is structured.
• How batch inference loads features and generates predictions.
• Where predictions are stored and how to inspect them.

Before you start

• Training is complete and a model is registered (Tutorial 5).

Step 1: Deploy the inference DAG

make -C projects/pudo deploy-inference-dag

This creates a Snowflake task graph that orchestrates the inference pipeline:

1. Load model: retrieves the latest (or configured) model version from the Model Registry.
2. Generate features: computes inference-time features from the feature store for the target date.
3. Run predictions: applies the model to the feature matrix.
4. Write results: stores predictions in the project schema.

Step 2: Run the inference DAG

make -C projects/pudo run-inference-dag

This triggers an immediate execution of the inference task graph.

Alternative: CLI-based inference

You can also run inference directly without the DAG:

make -C projects/pudo run-inference

This runs pudo-inference run, which performs the same steps as the DAG but from your local terminal. This is useful for debugging or ad-hoc runs.

How inference works

The inference pipeline:

1. Reads the model version from configuration (or uses the latest).
2. Constructs an inference spine, one row per (PUDO, target_date).
3. Performs ASOF JOINs against feature views to get point-in-time features.
4. Applies the trained XGBoost model to generate predictions.
5. Writes predictions with metadata (model version, run timestamp, features used) to the project schema.

Step 3: Inspect predictions

-- View recent predictions
SELECT *
FROM PUDO_DEV.PREDICTIONS
ORDER BY PREDICTION_DATE DESC
LIMIT 20;

-- Check prediction distribution
SELECT
  PREDICTION_DATE,
  COUNT(*) AS num_predictions,
  AVG(PREDICTED_CAPACITY) AS avg_predicted,
  STDDEV(PREDICTED_CAPACITY) AS std_predicted
FROM PUDO_DEV.PREDICTIONS
GROUP BY PREDICTION_DATE
ORDER BY PREDICTION_DATE DESC;

Inference CLI verbs

The pudo-inference CLI provides additional verbs for post-inference operations:

Command	What it does
pudo-inference run	Run batch inference.
pudo-inference evaluate	Compare predictions to actuals and compute metrics.
pudo-inference alerts	Check for alert conditions (e.g., high prediction error).
pudo-inference summary	Print a summary of recent predictions.

You will use evaluate, alerts, and summary in Tutorial 7.

What you have now

• Inference DAG deployed and executed.
• Batch predictions generated and stored in Snowflake.
• Understanding of the inference CLI verbs.

Next step

Continue to Tutorial 7: Simulate, Evaluate & Alert to simulate daily data cycles, evaluate prediction quality, and trigger alerts.