Skip to content

Skill: Zero-Copy Pipeline Generator (Coordinator)

This is the coordinator skill that orchestrates the end-to-end creation of Spark-centric and BigQuery-centric data pipelines for a new use case. It manages the user interview, schema investigation, architecture recommendation, and delegates code generation to specialized sub-skills.

Workflow

1. User Interview

Prompt the user to understand the new use case. Ask:

  1. Business Goal: What is the prediction target? (e.g., churn, recommendation).
  2. Datasets: What are the input BigQuery tables?
  3. Features & Label: What are the features and the target label?
  4. Data Prep Logic: How should the data be prepared/joined?
  5. Model Complexity: Do they need a simple model (like Logistic Regression) or a complex one (like Random Forest or XGBoost)?

2. Schema Investigation

Use the bq CLI tool to inspect the schemas of the tables provided by the user.

  • For each table, run:

    bq show --format=prettyjson <project_id>:<dataset_id>.<table_id>

  • Analyze the schema to confirm column names, types, and nullability. Present a summary to the user.

3. Solution Recommendation

Based on the interview and schema, propose the architecture:

  1. Pipeline Options:
    • Spark-Centric: Recommend if the user has existing Spark workloads, needs custom PySpark logic, or prefers Spark MLlib.
    • BigQuery-Centric: Recommend if the user wants a NoOps, SQL-backed pipeline using BigQuery DataFrames.
  2. Serving Options:
    • Lean JSON: Recommend if using Spark + Simple Model (Logistic Regression).
    • ONNX: Recommend if using Spark + Complex Model (Random Forest/XGBoost).
    • Native BQML: Recommend if using BigQuery-Centric pipeline.

Get the user's approval on the recommendation before proceeding.

4. Orchestrating Sub-Skills for Code Generation

Once approved, generate the notebooks by activating the relevant sub-skills:

Spark-Centric Pipeline Generation

If generating the Spark-centric notebook (spark_centric_<use_case>.ipynb), apply the following sub-skills:

  1. spark-lightning-optimization: Use to generate the Dataproc Serverless session with Lightning Engine.
  2. zero-copy-ingestion: Use to generate the BigQuery data loading code using the connector.
  3. model-evaluation: Use to generate the evaluation section (Accuracy or AUC-PR/Confusion Matrix).
  4. unified-model-registry: Use to generate the model saving and serving code (Lean JSON or ONNX export, Vertex AI registration).

BigQuery-Centric Pipeline Generation

If generating the BigQuery-centric notebook (bq_centric_<use_case>.ipynb), apply the following sub-skills:

  1. zero-copy-ingestion: Use to generate the BigFrames data loading code.
  2. bigframes-bqml: Use to generate the feature engineering and BQML training code.
  3. model-evaluation: Use to generate the evaluation section.
  4. unified-model-registry: Use to generate the BQML export and Vertex AI Endpoint deployment code.

Batch Prediction Script

Generate predict_job_<use_case>.py (PySpark) if the Spark pipeline is chosen, ensuring it loads the model from GCS and writes predictions back to BigQuery.