ETL (extract, transform, load) is a data integration process that copies data from several source systems, cleans and reshapes it, and loads it into a target store such as a data warehouse. Developers then build applications on that prepared data, and business users rely on it for decisions.
This page is for data engineers, analysts and data leaders who need a plain definition of ETL. It covers the three steps, how ETL differs from ELT, where it is used, and what to settle first.
The IBM explanation calls ETL a data integration process that combines, cleans and organises data from multiple sources. AWS describes it as combining data from multiple sources into a central repository called a data warehouse, with business logic applied so the raw data is ready for storage, analysis and machine learning.
In practice ETL does three jobs: it extracts data from the source systems, enforces data quality and consistency standards, and delivers data in a format that people and applications can use. For the wider discipline behind it, see our page on business intelligence.
Per IBM, the extract step copies or exports raw data from sources such as databases, CRM and ERP systems, IoT devices, APIs and files into a staging area. In the transform step, business rules clean, deduplicate, validate, calculate and reformat the data to match the target schema, and sensitive data can be masked or encrypted. The load step moves the prepared data into the warehouse. Traditional ETL does an initial full load followed by periodic incremental loads, while streaming ETL loads data continuously.
| Step | What happens | Typical output |
|---|---|---|
| Extract | Raw data is copied from source systems to a staging area | Unchanged source records |
| Transform | Rules clean, deduplicate, validate and reshape the data | Data that fits the target schema |
| Load | Prepared data moves into the target warehouse | Analysis-ready tables |
ELT reverses the order: data is loaded into the target system first and transformed afterwards. AWS says ETL requires more definition at the beginning, because teams must define target data types, structures and relationships upfront. IBM adds that ETL suits smaller, structured datasets, while ELT is designed for big data and handles structured, semi-structured and unstructured data.
| Question | ETL | ELT |
|---|---|---|
| When is data transformed? | Before loading, in a staging area | After loading, inside the target system |
| Planning | More definition needed upfront | Less upfront definition |
| Data types | Best with structured data in a predefined schema | Structured, semi-structured and unstructured |
| Typical fit | Smaller datasets, masking sensitive fields before delivery | High-volume data that loads frequently |
IBM lists several uses. PII can be masked during transformation before data reaches its destination. Batch ETL aggregates data for scheduled sales and compliance reports. Streaming ETL lets financial institutions spot suspicious activity and flag anomalies quickly, and processing IoT sensor data helps teams plan repairs and prevent equipment downtime. The shared benefit is that downstream users work from consistent, checked data instead of raw extracts.
An AI agent acts on whatever the pipeline delivers, so a transformation rule that changes a field also changes what the agent decides. That makes the rules, owners and lineage of each pipeline part of agent governance. ElixirData is XenonStack's Context OS, and it is where this governed context is considered alongside the data it comes from.
To map one decision your agents make, and the pipeline data and policy it depends on, book a Context Workshop. See what happens in the workshop.
ETL stands for extract, transform and load: data is extracted from source systems, transformed to fit the target schema, and loaded into a target such as a data warehouse.
In ETL, data is transformed before it is loaded. In ELT, data is loaded first and transformed inside the target system, which suits large volumes and mixed data types.
A staging area is a temporary location where extracted data is held while it is cleaned, deduplicated and reshaped before loading.
Batch ETL loads data in an initial full load and then periodic incremental loads, while streaming ETL loads data continuously.
Because transformation happens before delivery, personally identifiable information can be masked or encrypted before it reaches the target destination.