ETL and ELT are the two orders in which data is copied from the systems where it is created into a data warehouse. ETL stands for extract, transform, load. ELT stands for extract, load, transform. The tools that do this copying are also called data integration tools or data pipeline tools.
Each letter is one step. Extract means pulling the data out of the source system, the system where the data was first created, such as an order database or a CRM. Transform means cleaning the data, fixing formats and joining tables together. Load means putting the data into the data warehouse.
The older order is ETL. The data was cleaned on a separate server before it reached the warehouse, because warehouse storage and computing were expensive. Only the finished tables were loaded. ELT reverses the last two steps. Cloud warehouses made storage cheap, so the raw data is loaded first, unchanged. It is then cleaned inside the warehouse, with SQL, in a step called data transformation. Because the raw copy is kept, a mistake in the cleaning can be fixed and run again without extracting everything a second time.
An ELT product makes the extract and load steps ready to use. It offers hundreds of data connectors, ready-made code that knows how to pull data out of one particular app or database. Most connectors use each app's API, the set of requests that the app accepts from other programs, such as "send me all bookings changed since 10 a.m." The product runs each connector on a schedule, copies only what has changed since the last run, and adjusts the warehouse tables when a source adds a column.
Take, for example, a chain of diagnostic labs in Chennai with 120 sample collection centres. Its bills sit in a relational database. Its test results sit in the lab system. Its home sample bookings sit in a CRM, and its online payments come from a payment gateway. Every hour, the ELT tool asks each system for its new and changed records, such as 3,400 tests booked that morning. It loads them, raw, into the warehouse. A transformation then joins the four sources into one table of tests by centre and by day. The operations manager can now see which centres deliver reports late.
Some products are no-code: the data team picks a source and a destination, and the tool does the rest. Others offer a visual builder, in which engineers draw the pipeline and the tool writes the code, which they can read and change. The newest generate the pipeline from a written description. Some data teams buy one suite that also does the cataloguing and checking, while others join separate tools, as the page on the modern data stack explains.
How is ETL and ELT different from change data capture? An ELT tool usually copies on a schedule, so the warehouse can be an hour or a night behind; the page on batch and stream processing compares the two. And how is it different from an integration platform (iPaaS)? An iPaaS connects applications so that an action in one starts an action in another. ETL and ELT copy data for analysis. Data observability checks each run for missing or late data.