NeuralBlitz ETL Pipelines
What I do
I provide frameworks and patterns for building robust data pipelines that extract data from various sources, transform it to meet business requirements, and load it into target systems. I ensure data quality, enable scheduled execution, and provide monitoring and error handling capabilities for production data workflows.
When to use me
- Moving data from operational databases to data warehouses
- Consolidating data from multiple sources
- Creating data lakes from raw data
- Building feature stores for machine learning
- Scheduling regular data refreshes
- Implementing data quality checks
- Building real-time streaming pipelines
- Creating data aggregation workflows
Core Concepts
Pipeline Components
- Extract: Pull data from source systems (databases, APIs, files)
- Transform: Clean, validate, enrich, aggregate data
- Load: Write data to destination systems
- Orchestrate: Manage dependencies and scheduling
Data Quality
- Schema validation: Verify structure matches expectations
- Data type checks: Ensure correct types
- Range checks: Validate value ranges
- Uniqueness checks: Detect duplicate records
- Referential integrity: Verify relationships
Orchestration
- DAGs: Directed Acyclic Graphs for dependency management
- Task dependencies: Define execution order
- Failure handling: Retries, alerts, dead letter queues
- Backfilling: Re-processing historical data
Tools
- Apache Airflow: Workflow orchestration
- Prefect/Dagster: Modern orchestration frameworks
- Luigi: Spotify's pipeline framework
- dbt: Data transformation tool
Code Examples (Python)
import pandas as pd
import numpy as np
from datetime import datetime, timedelta
import logging
# Basic ETL implementation
class BasicETL:
def __init__(self, source_conn, target_conn):
self.source = source_conn
self.target = target_conn
self.logger = logging.getLogger(__name__)
def extract(self, query, params=None):
"""Extract data from source."""
self.logger.info(f"Extracting data with query: {query[:100]}...")
df = pd.read_sql(query, self.source, params=params)
self.logger.info(f"Extracted {len(df)} rows")
return df
def transform(self, df):
"""Transform data."""
self.logger.info(f"Transforming {len(df)} rows...")
# Clean column names
df.columns = [c.strip().lower().replace(' ', '_') for c in df.columns]
# Handle missing values
df = df.fillna({
'numeric_col': 0,
'string_col': 'unknown',
'date_col': pd.NaT
})
# Type conversion
df['date_col'] = pd.to_datetime(df['date_col'], errors='coerce')
df['numeric_col'] = pd.to_numeric(df['numeric_col'], errors='coerce')
# Remove duplicates
df = df.drop_duplicates(subset=['id'])
# Derived columns
df['processed_at'] = datetime.now()
# Filter invalid records
df = df[df['value'] >= 0]
return df
def load(self, df, table_name, if_exists='append'):
"""Load data to target."""
self.logger.info(f"Loading {len(df)} rows to {table_name}...")
df.to_sql(table_name, self.target, if_exists=if_exists, index=False)
self.logger.info(f"Successfully loaded {len(df)} rows")
def run(self, extract_query, target_table, transform_func=None):
"""Execute full ETL pipeline."""
try:
df = self.extract(extract_query)
df = self.transform(df)
if transform_func:
df = transform_func(df)
self.load(df, target_table)
return True
except Exception as e:
self.logger.error(f"ETL failed: {e}")
raise
# Data validation
class DataValidator:
@staticmethod
def validate_schema(df, expected_schema):
"""Validate DataFrame schema."""
for col, dtype in expected_schema.items():
if col not in df.columns:
raise ValueError(f"Missing column: {col}")
if df[col].dtype != dtype:
raise TypeError(f"Column {col} has wrong type: {df[col].dtype} vs {dtype}")
@staticmethod
def validate_range(df, col, min_val=None, max_val=None):
"""Validate values in range."""
if min_val is not None and (df[col] < min_val).any():
invalid = (df[col] < min_val).sum()
raise ValueError(f"{col} has {invalid} values below {min_val}")
if max_val is not None and (df[col] > max_val).any():
invalid = (df[col] > max_val).sum()
raise ValueError(f"{col} has {invalid} values above {max_val}")
@staticmethod
def validate_uniqueness(df, col):
"""Check for duplicates."""
duplicates = df[col].duplicated().sum()
if duplicates > 0:
raise ValueError(f"Column {col} has {duplicates} duplicates")
@staticmethod
def validate_completeness(df, required_cols, threshold=0.99):
"""Check completeness of required columns."""
for col in required_cols:
completeness = df[col].notna().mean()
if completeness < threshold:
raise ValueError(f"Column {col} has only {completeness:.2%} completeness")
# Incremental ETL
class IncrementalETL:
def __init__(self, source, target):
self.source = source
self.target = target
def get_last_processed(self, table_name):
"""Get timestamp of last successful load."""
try:
query = f"SELECT MAX(processed_at) as last_run FROM etl_log WHERE table_name = '{table_name}'"
result = pd.read_sql(query, self.target)
return result['last_run'].iloc[0] if len(result) > 0 else None
except:
return None
def log_etl(self, table_name, rows_processed, status):
"""Log ETL execution."""
log_df = pd.DataFrame([{
'table_name': table_name,
'rows_processed': rows_processed,
'status': status,
'run_time': datetime.now()
}])
log_df.to_sql('etl_log', self.target, if_exists='append', index=False)
def run_incremental(self, table_name, watermark_col, extract_query_template):
"""Run incremental ETL based on watermark."""
last_run = self.get_last_processed(table_name)
if last_run:
query = extract_query_template.format(watermark=f"WHERE {watermark_col} > '{last_run}'")
else:
query = extract_query_template.format(watermark="")
df = pd.read_sql(query, self.source)
if len(df) == 0:
self.logger.info("No new data to process")
return
# Transform
df = self.transform(df)
# Load
self.load(df, table_name, if_exists='append')
# Log success
self.log_etl(table_name, len(df), 'success')
return df
# Airflow DAG example
"""
from airflow import DAG
from airflow.operators.python import PythonOperator
from airflow.operators.postgres_operator import PostgresOperator
from datetime import datetime, timedelta
default_args = {
'owner': 'data_team',
'depends_on_past': False,
'start_date': datetime(2024, 1, 1),
'email_on_failure': True,
'email_on_retry': False,
'retries': 3,
'retry_delay': timedelta(minutes=5),
}
dag = DAG(
'daily_etl_pipeline',
default_args=default_args,
description='Daily ETL pipeline for data warehouse',
schedule_interval='0 2 * * *', # Run at 2 AM daily
catchup=False,
)
def extract_sales(**context):
from sqlalchemy import create_engine
engine = create_engine(context['connection']['sales_db'])
df = pd.read_sql("SELECT * FROM sales", engine)
df.to_csv('/tmp/sales_raw.csv', index=False)
return '/tmp/sales_raw.csv'
def transform_sales(**context):
df = pd.read_csv(context['task_instance'].xcom_pull(task_ids='extract_sales'))
df['date'] = pd.to_datetime(df['date'])
df['revenue'] = df['quantity'] * df['price']
df.to_parquet('/tmp/sales_transformed.parquet', index=False)
return '/tmp/sales_transformed.parquet'
def load_warehouse(**context):
df = pd.read_parquet(context['task_instance'].xcom_pull(task_ids='transform_sales'))
engine = create_engine(context['connection']['warehouse'])
df.to_sql('daily_sales', engine, if_exists='append', index=False)
extract_task = PythonOperator(
task_id='extract_sales',
python_callable=extract_sales,
provide_context=True,
dag=dag,
)
transform_task = PythonOperator(
task_id='transform_sales',
python_callable=transform_sales,
provide_context=True,
dag=dag,
)
load_task = PythonOperator(
task_id='load_warehouse',
python_callable=load_warehouse,
provide_context=True,
dag=dag,
)
extract_task >> transform_task >> load_task
"""
# Streaming ETL with Kafka
"""
from kafka import KafkaConsumer, KafkaProducer
import json
class StreamingETL:
def __init__(self, kafka_brokers, input_topic, output_topic):
self.consumer = KafkaConsumer(
input_topic,
bootstrap_servers=kafka_brokers,
value_deserializer=lambda x: json.loads(x.decode('utf-8'))
)
self.producer = KafkaProducer(
bootstrap_servers=kafka_brokers,
value_serializer=lambda x: json.dumps(x).encode('utf-8')
)
def transform_record(self, record):
"""Transform single record."""
return {
'id': record['id'],
'timestamp': datetime.now().isoformat(),
'value_normalized': (record['value'] - record['mean']) / record['std'],
'category': record['category'].upper()
}
def run(self):
"""Process streaming records."""
for message in self.consumer:
record = message.value
transformed = self.transform_record(record)
self.producer.send(output_topic, value=transformed)
"""
Best Practices
Design for failure: Assume any component can fail; implement retries, logging, and alerts.
Incremental over full reload: Process only new/changed data when possible for efficiency.
Idempotent operations: Running the pipeline multiple times should produce the same result.
Separate concerns: Keep extraction, transformation, and loading logic independent.
Metadata management: Track data lineage, schema versions, and processing timestamps.
Monitor data quality: Implement validation at each stage with alerting.
Handle schema evolution: Plan for schema changes without breaking pipelines.
Document data contracts: Define clear interfaces between pipeline stages.
Common Patterns
Pattern 1: Resilient Batch ETL
class ResilientETL:
def __init__(self, config):
self.config = config
self.max_retries = 3
self.backoff_factor = 2
def run_with_retry(self, task_name, task_func, **kwargs):
"""Run task with exponential backoff retry."""
for attempt in range(self.max_retries):
try:
return task_func(**kwargs)
except Exception as e:
if attempt == self.max_retries - 1:
self.send_alert(task_name, e)
raise
sleep_time = self.backoff_factor ** attempt
self.logger.warning(f"Task {task_name} failed, retrying in {sleep_time}s")
sleep(sleep_time)
def checkpoint(self, stage, state):
"""Save pipeline state for recovery."""
checkpoint_df = pd.DataFrame([{
'stage': stage,
'state': json.dumps(state),
'timestamp': datetime.now()
}])
checkpoint_df.to_sql('pipeline_checkpoints',
self.checkpoint_db,
if_exists='append')
def resume_from_checkpoint(self, stage):
"""Resume pipeline from last checkpoint."""
query = f"SELECT * FROM pipeline_checkpoints WHERE stage = '{stage}' ORDER BY timestamp DESC LIMIT 1"
return pd.read_sql(query, self.checkpoint_db).iloc[0]['state']
Pattern 2: Data Quality Framework
class DataQualityFramework:
def __init__(self):
self.checks = []
def add_check(self, name, check_func, severity='error'):
"""Add quality check."""
self.checks.append({'name': name, 'func': check_func, 'severity': severity})
def run_checks(self, df, context):
"""Execute all quality checks."""
results = []
for check in self.checks:
try:
passed = check['func'](df, context)
results.append({
'check': check['name'],
'passed': passed,
'severity': check['severity']
})
except Exception as e:
results.append({
'check': check['name'],
'passed': False,
'error': str(e),
'severity': check['severity']
})
# Raise on error severity failures
errors = [r for r in results if not r['passed'] and r['severity'] == 'error']
if errors:
raise DataQualityError(errors)
return results
@staticmethod
def standard_checks():
"""Create standard quality checks."""
framework = DataQualityFramework()
framework.add_check('null_check',
lambda df, ctx: df.isnull().sum().sum() == 0)
framework.add_check('schema_check',
lambda df, ctx: set(df.columns) == set(ctx['expected_columns']))
framework.add_check('range_check',
lambda df, ctx: (df[ctx['numeric_cols']] >= 0).all().all())
framework.add_check('duplicate_check',
lambda df, ctx: df.duplicated().sum() == 0)
return framework
Pattern 3: Feature Engineering Pipeline
class FeatureEngineeringPipeline:
def __init__(self, feature_store):
self.feature_store = feature_store
self.transformations = []
def add_feature(self, name, source_col, transform_func):
"""Add feature transformation."""
self.transformations.append({
'name': name,
'source': source_col,
'func': transform_func
})
def build_features(self, df, timestamp_col='timestamp'):
"""Build all features for a dataset."""
features_df = df.copy()
for feat in self.transformations:
features_df[feat['name']] = feat['func'](features_df[feat['source']])
# Add temporal features
features_df['hour'] = features_df[timestamp_col].dt.hour
features_df['day_of_week'] = features_df[timestamp_col].dt.dayofweek
features_df['is_weekend'] = features_df['day_of_week'].isin([5, 6]).astype(int)
# Add rolling features
features_df = self.add_rolling_features(features_df, 'value', '1h', '24h')
return features_df
def add_rolling_features(self, df, col, *windows):
"""Add rolling window features."""
for window in windows:
df[f'{col}_rolling_{window}_mean'] = df.groupby('id')[col].transform(
lambda x: x.rolling(window, min_periods=1).mean()
)
return df
def save_to_store(self, entity_id, features_df, timestamp):
"""Save features to feature store."""
self.feature_store.save(
entity_id=entity_id,
features=features_df.to_dict(orient='records'),
timestamp=timestamp
)