ELT: Load Raw Data, Transform in Place

WHY IT EXISTS ELT exists to leverage the massive, scalable compute power of modern cloud data warehouses. Instead of requiring a separate, powerful transformation engine before data even lands (like in ETL), ELT uses the destination system's own resources. This simplifies the data ingestion pipeline by pushing the complex work to the system designed to handle big data.

THE MENTAL MODEL Think of it like unpacking groceries. The traditional ETL way is to sort and prep everything on your kitchen counter (a staging area) before putting it in the fridge (the warehouse). The ELT way is to dump all the grocery bags directly into a massive, smart fridge that sorts, cleans, and organizes everything for you internally, on demand.

HOW IT WORKS The process has three steps in a specific order. First, data is Extracted from various sources like databases or APIs. Second, this raw, unaltered data is Loaded directly into a target system, typically a cloud data warehouse or feature store. Third, Transformation logic, usually written in SQL, is executed directly within the warehouse to clean, standardize, and prepare the data for analysis or machine learning.

WHEN TO USE IT ELT shines when you're using a powerful cloud data warehouse (like Snowflake, BigQuery, or Redshift) that can handle large-scale transformations efficiently. It's ideal when you want to store the raw data for future, unforeseen analyses and for building ML feature pipelines where you refine raw data into features inside a feature store or warehouse.

WHEN NOT TO USE IT If your destination system has limited compute power, running complex transformations inside it could create performance bottlenecks. Traditional ETL is often better when data needs heavy cleaning, masking, or complex enrichment before it's exposed to the destination system, perhaps for compliance or security reasons. ELT couples your transformations to the warehouse; ETL decouples them.

ONE CANONICAL EXAMPLE A feature pipeline for an ML model. Raw user activity logs are extracted from production systems and loaded as-is into a data warehouse. A scheduled SQL query then runs inside the warehouse to aggregate this raw data, calculating metrics like 'user_session_duration' or 'purchase_count_last_30_days'. These calculated columns are the new features, now ready for the ML model to use, created without a separate transformation service.