data-engineering

Data Engineering

Pipeline Architecture

ETL vs ELT

Pattern	When to Use	Tools
ETL	Transform before loading, data quality critical	Airflow + custom, Spark
ELT	Raw → warehouse → transform in-place	Fivetran + dbt, Airbyte + dbt

Orchestration

Apache Airflow:

from airflow.decorators import dag, task
from datetime import datetime

@dag(schedule="@daily", start_date=datetime(2024, 1, 1), catchup=False)
def my_pipeline():
    @task()
    def extract() -> dict:
        return {"data": "extracted"}

    @task()
    def transform(data: dict) -> dict:
        return {"transformed": True}

    @task()
    def load(data: dict):
        # Load to warehouse
        pass

    raw = extract()
    transformed = transform(raw)
    load(transformed)

my_pipeline()

Dagster (recommended for new projects):

from dagster import asset, Definitions

@asset
def raw_users():
    return extract_from_source()

@asset
def cleaned_users(raw_users):
    return clean_and_validate(raw_users)

dbt Transformations

-- models/marts/dim_customers.sql
{{ config(materialized='table', schema='marts') }}

WITH source AS (
    SELECT * FROM {{ ref('stg_customers') }}
),
orders AS (
    SELECT customer_id, COUNT(*) as order_count, SUM(amount) as total_spent
    FROM {{ ref('stg_orders') }}
    GROUP BY customer_id
)
SELECT
    s.customer_id,
    s.name,
    s.email,
    COALESCE(o.order_count, 0) as lifetime_orders,
    COALESCE(o.total_spent, 0) as lifetime_value
FROM source s
LEFT JOIN orders o ON s.customer_id = o.customer_id

Stream Processing

Apache Kafka:

from confluent_kafka import Producer, Consumer

# Producer
producer = Producer({'bootstrap.servers': 'localhost:9092'})
producer.produce('events', key='user_123', value=json.dumps(event))
producer.flush()

# Consumer
consumer = Consumer({
    'bootstrap.servers': 'localhost:9092',
    'group.id': 'my-group',
    'auto.offset.reset': 'earliest'
})
consumer.subscribe(['events'])

Data Warehouse Schema Design

Star Schema

• Fact tables: Measurable events (orders, clicks, transactions)
• Dimension tables: Descriptive context (customers, products, dates)
• Slowly Changing Dimensions: Type 1 (overwrite), Type 2 (versioned rows), Type 3 (previous column)

Data Quality

• Great Expectations: Schema validation, statistical tests, custom expectations
• dbt tests: not_null, unique, accepted_values, relationships, custom SQL tests
• Data contracts: Schema evolution policies, backward compatibility requirements

Key Patterns

• Idempotent pipelines: Same input always produces same output, safe to rerun
• Incremental models: Process only new/changed data, use updated_at watermarks
• Dead letter queues: Route failed records for inspection without blocking pipeline
• Backfill strategy: Time-partitioned tables enable targeted historical reprocessing