What is ELT, and how is it different from ETL?

ETL and ELT use the same 3 letters: extract, load, transform. The only difference is the order, and that order says a lot about what storage and computing cost when a system was designed.

ETL: clean it before you store it

In ETL, data is extracted from a source, transformed on the way (cleaned, joined, summed) and only then loaded into the warehouse. For decades this made sense. Warehouse storage and computing were expensive and often ran on a company's own hardware, so you stored only the finished result, in exactly the shape the reports needed.

The weakness is that the raw data never arrives. If the business later wants a column the transformation dropped, or finds a bug in how refunds were handled, you have to go back to the source and run it all again, if the source still has the data.

ELT: store it raw, change it in place

ELT flips the last 2 steps. Raw data is loaded first, as it came, and transformed inside the warehouse or lakehouse with that system's own computing power, usually in SQL.

This became the common order with cloud platforms, which made storage and computing available on demand, at scale, paid for as used. Once storing raw data is cheap and the warehouse can transform it quickly, there is little reason to throw anything away before it lands.

The main gain is the one ETL lacked: the raw data is kept, so a transformation can be fixed or changed later and rerun over everything. Tools such as dbt are built for that T step.

ELT in a lakehouse

In a LakehouseLakehouseOne store for all of an organisation's data: cheap files in cloud storage, as in a data lake, with the tables, transactions and fast queries of a data warehouse on top., ELT is the medallion architectureMedallion ArchitectureOrganising a lakehouse in 3 layers of rising quality: bronze for raw data, silver for cleaned and joined data, and gold for tables shaped for reports and AI.. The bronze layer is the L: data as it arrived, often with most fields kept as plain strings, with full history so it can be processed again. Silver and gold are the T, done step by step inside the platform.

The real 2-tier setups of the last decade were a mix. Data was extracted and loaded into a lakeData LakeCheap storage that holds data of any kind in its raw form, usually as files in cloud object storage, to be cleaned and structured later., then transformed and loaded again into a warehouseData WarehouseA database built for analysis: cleaned, structured data from many systems, kept in one place and shaped for fast reports and dashboards.. Each extra copy was another job that could fail and another place for the numbers to drift apart.

The cost of keeping everything

ELT is not free. Raw data includes what ETL would have filtered out: personal details, fields nobody should read, test records. Keeping it means governing it, with access rules on the raw layer as well as the finished one. And the transformations still need checks, because a bad transform in the warehouse is just as wrong as a bad one on the way in.

Why it matters for AI

AI work changes its mind often. A better embedding model arrives, or a different way of chunkingChunkingCutting documents into smaller passages before they are embedded for retrieval, so a search returns the part that answers a question rather than a whole file. documents works better. On Databricks, switching the embedding model of a self-managed AI Search index means a new index and new embeddings for every document. If the raw text was kept, that is a rerun. If only the old vectors were kept, it is a rebuild from the sources, if they are still there. That is the ELT argument again, in a new place.