ml-pipeline-automation（毫升管道自动化）

ML Pipeline Automation

Orchestrate end-to-end machine learning workflows from data ingestion to production deployment with production-tested Airflow, Kubeflow, and MLflow patterns.

When to Use This Skill

Load this skill when:

• Building ML Pipelines: Orchestrating data → train → deploy workflows
• Scheduling Retraining: Setting up automated model retraining schedules
• Experiment Tracking: Tracking experiments, parameters, metrics across runs
• MLOps Implementation: Building reproducible, monitored ML infrastructure
• Workflow Orchestration: Managing complex multi-step ML workflows
• Model Registry: Managing model versions and deployment lifecycle

Quick Start: ML Pipeline in 5 Steps

# 1. Install Airflow and MLflow (check for latest versions at time of use)
pip install apache-airflow==3.1.5 mlflow==3.7.0

# Note: These versions are current as of December 2025
# Check PyPI for latest stable releases: https://pypi.org/project/apache-airflow/

# 2. Initialize Airflow database
airflow db init

# 3. Create DAG file: dags/ml_training_pipeline.py
cat > dags/ml_training_pipeline.py << 'EOF'
from airflow import DAG
from airflow.operators.python import PythonOperator
from datetime import datetime, timedelta

default_args = {
    'owner': 'ml-team',
    'retries': 2,
    'retry_delay': timedelta(minutes=5)
}

dag = DAG(
    'ml_training_pipeline',
    default_args=default_args,
    schedule_interval='@daily',
    start_date=datetime(2025, 1, 1)
)

def train_model(**context):
    import mlflow
    import mlflow.sklearn
    from sklearn.ensemble import RandomForestClassifier
    from sklearn.datasets import load_iris
    from sklearn.model_selection import train_test_split

    X, y = load_iris(return_X_y=True)
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

    mlflow.set_tracking_uri('http://localhost:5000')
    mlflow.set_experiment('iris-training')

    with mlflow.start_run():
        model = RandomForestClassifier(n_estimators=100)
        model.fit(X_train, y_train)

        accuracy = model.score(X_test, y_test)
        mlflow.log_metric('accuracy', accuracy)
        mlflow.sklearn.log_model(model, 'model')

train = PythonOperator(
    task_id='train_model',
    python_callable=train_model,
    dag=dag
)
EOF

# 4. Start Airflow scheduler and webserver
airflow scheduler &
airflow webserver --port 8080 &

# 5. Trigger pipeline
airflow dags trigger ml_training_pipeline

# Access UI: http://localhost:8080

Result: Working ML pipeline with experiment tracking in under 5 minutes.

Core Concepts

Pipeline Stages

1. Data Collection → Fetch raw data from sources
2. Data Validation → Check schema, quality, distributions
3. Feature Engineering → Transform raw data to features
4. Model Training → Train with hyperparameter tuning
5. Model Evaluation → Validate performance on test set
6. Model Deployment → Push to production if metrics pass
7. Monitoring → Track drift, performance in production

Orchestration Tools Comparison

Basic Airflow DAG

from airflow import DAG
from airflow.operators.python import PythonOperator
from datetime import datetime, timedelta
import logging

logger = logging.getLogger(__name__)

default_args = {
    'owner': 'ml-team',
    'depends_on_past': False,
    'email': ['alerts@example.com'],
    'email_on_failure': True,
    'retries': 2,
    'retry_delay': timedelta(minutes=5)
}

dag = DAG(
    'ml_training_pipeline',
    default_args=default_args,
    description='End-to-end ML training pipeline',
    schedule_interval='@daily',
    start_date=datetime(2025, 1, 1),
    catchup=False
)

def validate_data(**context):
    """Validate input data quality."""
    import pandas as pd

    data_path = "/data/raw/latest.csv"
    df = pd.read_csv(data_path)

    # Validation checks
    assert len(df) > 1000, f"Insufficient data: {len(df)} rows"
    assert df.isnull().sum().sum() < len(df) * 0.1, "Too many nulls"

    context['ti'].xcom_push(key='data_path', value=data_path)
    logger.info(f"Data validation passed: {len(df)} rows")

def train_model(**context):
    """Train ML model with MLflow tracking."""
    import mlflow
    import mlflow.sklearn
    from sklearn.ensemble import RandomForestClassifier

    data_path = context['ti'].xcom_pull(key='data_path', task_ids='validate_data')

    mlflow.set_tracking_uri('http://mlflow:5000')
    mlflow.set_experiment('production-training')

    with mlflow.start_run():
        # Training logic here
        model = RandomForestClassifier(n_estimators=100)
        # model.fit(X, y) ...

        mlflow.log_param('n_estimators', 100)
        mlflow.sklearn.log_model(model, 'model')

validate = PythonOperator(
    task_id='validate_data',
    python_callable=validate_data,
    dag=dag
)

train = PythonOperator(
    task_id='train_model',
    python_callable=train_model,
    dag=dag
)

validate >> train

Known Issues Prevention

1. Task Failures Without Alerts

Problem: Pipeline fails silently, no one notices until users complain.

Solution: Configure email/Slack alerts on failure:

default_args = {
    'email': ['ml-team@example.com'],
    'email_on_failure': True,
    'email_on_retry': False
}

def on_failure_callback(context):
    """Send Slack alert on failure."""
    from airflow.providers.slack.operators.slack_webhook import SlackWebhookOperator

    slack_msg = f"""
    :red_circle: Task Failed: {context['task_instance'].task_id}
    DAG: {context['task_instance'].dag_id}
    Execution Date: {context['ds']}
    Error: {context.get('exception')}
    """

    SlackWebhookOperator(
        task_id='slack_alert',
        slack_webhook_conn_id='slack_webhook',
        message=slack_msg
    ).execute(context)

task = PythonOperator(
    task_id='critical_task',
    python_callable=my_function,
    on_failure_callback=on_failure_callback,
    dag=dag
)

2. Missing XCom Data Between Tasks

Problem: Task expects XCom value from previous task, gets None, crashes.

Solution: Always validate XCom pulls:

def process_data(**context):
    data_path = context['ti'].xcom_pull(
        key='data_path',
        task_ids='upstream_task'
    )

    if data_path is None:
        raise ValueError("No data_path from upstream_task - check XCom push")

    # Process data...

3. DAG Not Appearing in UI

Problem: DAG file exists in dags/ but doesn't show in Airflow UI.

Solution: Check DAG parsing errors:

# Check for syntax errors
python dags/my_dag.py

# View DAG import errors in UI
# Navigate to: Browse → DAG Import Errors

# Common fixes:
# 1. Ensure DAG object is defined in file
# 2. Check for circular imports
# 3. Verify all dependencies installed
# 4. Fix syntax errors

4. Hardcoded Paths Break in Production

Problem: Paths like /Users/myname/data/ work locally, fail in production.

Solution: Use Airflow Variables or environment variables:

from airflow.models import Variable

def load_data(**context):
    # ❌ Bad: Hardcoded path
    # data_path = "/Users/myname/data/train.csv"

    # ✅ Good: Use Airflow Variable
    data_dir = Variable.get("data_directory", "/data")
    data_path = f"{data_dir}/train.csv"

    # Or use environment variable
    import os
    data_path = os.getenv("DATA_PATH", "/data/train.csv")

5. Stuck Tasks Consume Resources

Problem: Task hangs indefinitely, blocks worker slot, wastes resources.

Solution: Set execution_timeout on tasks:

from datetime import timedelta

task = PythonOperator(
    task_id='long_running_task',
    python_callable=my_function,
    execution_timeout=timedelta(hours=2),  # Kill after 2 hours
    dag=dag
)

6. No Data Validation = Bad Model Training

Problem: Train on corrupted/incomplete data, model performs poorly in production.

Solution: Add data quality validation tasks:

def validate_data_quality(**context):
    """Comprehensive data validation."""
    import pandas as pd

    df = pd.read_csv(data_path)

    # Schema validation
    required_cols = ['user_id', 'timestamp', 'feature_a', 'target']
    missing_cols = set(required_cols) - set(df.columns)
    if missing_cols:
        raise ValueError(f"Missing columns: {missing_cols}")

    # Statistical validation
    if df['target'].isnull().sum() > 0:
        raise ValueError("Target column contains nulls")

    if len(df) < 1000:
        raise ValueError(f"Insufficient data: {len(df)} rows")

    logger.info("✅ Data quality validation passed")

7. Untracked Experiments = Lost Knowledge

Problem: Can't reproduce results, don't know which hyperparameters worked.

Solution: Use MLflow for all experiments:

import mlflow

mlflow.set_tracking_uri('http://mlflow:5000')
mlflow.set_experiment('model-experiments')

with mlflow.start_run(run_name='rf_v1'):
    # Log ALL hyperparameters
    mlflow.log_params({
        'model_type': 'random_forest',
        'n_estimators': 100,
        'max_depth': 10,
        'random_state': 42
    })

    # Log ALL metrics
    mlflow.log_metrics({
        'train_accuracy': 0.95,
        'test_accuracy': 0.87,
        'f1_score': 0.89
    })

    # Log model
    mlflow.sklearn.log_model(model, 'model')

When to Load References

Load reference files for detailed production implementations:

• Airflow DAG Patterns: Load references/airflow-patterns.md when building complex DAGs with error handling, dynamic generation, sensors, task groups, or retry logic. Contains complete production DAG examples.
• Kubeflow & MLflow Integration: Load references/kubeflow-mlflow.md when using Kubeflow Pipelines for container-native orchestration, integrating MLflow tracking, building KFP components, or managing model registry.
• Pipeline Monitoring: Load references/pipeline-monitoring.md when implementing data quality checks, drift detection, alert configuration, or pipeline health monitoring with Prometheus.

Best Practices

1. Idempotent Tasks: Tasks should produce same result when re-run
2. Atomic Operations: Each task does one thing well
3. Version Everything: Data, code, models, dependencies
4. Comprehensive Logging: Log all important events with context
5. Error Handling: Fail fast with clear error messages
6. Monitoring: Track pipeline health, data quality, model drift
7. Testing: Test tasks independently before integrating
8. Documentation: Document DAG purpose, task dependencies

Common Patterns

Conditional Execution

from airflow.operators.python import BranchPythonOperator

def choose_branch(**context):
    accuracy = context['ti'].xcom_pull(key='accuracy', task_ids='evaluate')

    if accuracy > 0.9:
        return 'deploy_to_production'
    else:
        return 'retrain_with_more_data'

branch = BranchPythonOperator(
    task_id='check_accuracy',
    python_callable=choose_branch,
    dag=dag
)

train >> evaluate >> branch >> [deploy, retrain]

Parallel Training

from airflow.utils.task_group import TaskGroup

with TaskGroup('train_models', dag=dag) as train_group:
    train_rf = PythonOperator(task_id='train_rf', ...)
    train_lr = PythonOperator(task_id='train_lr', ...)
    train_xgb = PythonOperator(task_id='train_xgb', ...)

# All models train in parallel
preprocess >> train_group >> select_best

Waiting for Data

from airflow.sensors.filesystem import FileSensor

wait_for_data = FileSensor(
    task_id='wait_for_data',
    filepath='/data/input/{{ ds }}.csv',
    poke_interval=60,  # Check every 60 seconds
    timeout=3600,  # Timeout after 1 hour
    mode='reschedule',  # Don't block worker
    dag=dag
)

wait_for_data >> process_data