While the ETL process transforms data before loading it into the destination, there's another common pattern in data pipelines: Extract, Load, Transform (ELT). As the name suggests, ELT changes the order of operations. Instead of transforming data mid-flight, it loads the raw or minimally processed data directly into the target system first, and then performs transformations within that target system.

This approach became more popular with the rise of powerful, scalable cloud data warehouses and data lakes. These systems often have significant computational power capable of handling large-scale transformations efficiently.

Let's break down the ELT process:

Extract

This step is identical to the 'Extract' phase in ETL. Data is retrieved from its original sources. These sources can be diverse, including:

Relational databases (like PostgreSQL, MySQL)
NoSQL databases
Application Programming Interfaces (APIs)
Log files
Files from storage systems (like CSV, JSON, Parquet files)

The goal here is simply to get the data out of the source system.

Load

Here lies the main difference from ETL. In the ELT pattern, the extracted data is loaded almost immediately into the target storage system, typically a data lake or a data warehouse. Minimal cleaning or structuring might occur, but the heavy transformations are deferred.

For example, raw JSON data from an API might be loaded directly into a staging table or area within a data warehouse, or dropped as files into a data lake. The structure isn't necessarily enforced strictly at this stage. This allows for faster data ingestion because the pipeline doesn't wait for potentially time-consuming transformations.

Transform

Only after the data resides within the target system (the data warehouse or data lake) does the transformation step occur. Data engineers or analysts can then use the processing capabilities of the target system itself to clean, enrich, aggregate, join, and reshape the data into the desired format for analysis or application use.

Often, this transformation step is performed using SQL within a data warehouse, or using processing frameworks like Apache Spark that can operate directly on data within a data lake or warehouse.

A diagram illustrating the sequence of operations in an ELT pipeline: Extract data from sources, Load it into the target system, and then Transform it within that system.

Why Choose ELT?

The ELT approach offers several advantages, particularly in modern data environments:

Faster Ingestion: Since transformations don't happen before loading, data becomes available in the target system much quicker.
Flexibility: Raw data is stored in the target system. This means you can apply different transformations later for various purposes without having to re-extract the data. If analytical requirements change, new transformation logic can be applied to the existing raw data.
Leveraging Target System Power: Cloud data warehouses (like Google BigQuery, Amazon Redshift, Snowflake) and data lake query engines are designed to handle large-scale data processing efficiently. ELT takes advantage of this power for transformations.
Handling Diverse Data: ELT is well-suited for data lakes where you might store structured, semi-structured, and unstructured data. You load everything first and then decide how to process and structure it later (sometimes referred to as "schema-on-read").

ELT vs. ETL: The Key Difference

The fundamental distinction lies in when the transformation happens.

ETL: Extract -> Transform -> Load (Transformation occurs before loading into the final target).
ELT: Extract -> Load -> Transform (Transformation occurs after loading into the target).

ELT is often preferred when dealing with large data volumes, leveraging powerful cloud data platforms, and when flexibility in applying transformations is desired. You load the raw ingredients first, then decide on the recipe inside the kitchen (your data warehouse or lake).