ETL and ELT chain the same three operations, extract, transform, load, and only the order of the last two separates them. The question “ETL or ELT” therefore concerns a place of execution: data transformation runs either in an engine placed before the target, or in the target itself, once the raw data has been loaded.
We adopt four criteria, which tip the balance one way or the other: where the computing power lies, the sensitivity of the data, the team’s skills and the way compute is billed. When the target is a warehouse or a lakehouse whose compute scales with the load, and nothing prohibits depositing raw data there, ELT is the starting point. When the target is constrained, or certain data must not enter it as is, the transformation stays upstream. A single platform can combine both orders, flow by flow.
ETL or ELT, a question of where the work runs
The Microsoft and BigQuery documentation, read on 12 November 2026, frame the choice in these terms. The first (Azure Architecture Center, page “Extract, transform, and load (ETL)”) states that ELT differs from ETL only in where the transformation takes place, and that it only works well if the target system is powerful enough to transform the data efficiently. The second (page “Introduction to loading, transforming, and exporting data”) generally recommends ELT to most of its customers and considers ETL useful when a transformation process already exists, or to reduce resource consumption in the warehouse. Each speaks for its own products: we do not adopt their preference, only two variables, the capacity of the target for the first, what one is willing to consume there for the second.
The first criterion follows from this: the transformation is placed where compute is available without hindering readers. A warehouse or lakehouse whose compute is sized independently of storage absorbs transformations alongside analytical queries. ELT is practicable there, and the retained raw data makes it possible to replay a corrected rule without calling on the source again. The assembly of tools that has formed around ELT is described in our modern data stack entry.
The situation is reversed when the target is a database sized for reporting, an operational application or a fixed-capacity system, already loaded during working hours. Running heavy joins there degrades the service provided to readers, and the transformation is done in a separate engine. It is also done upstream when the useful compute lies close to the source, for example to filter or aggregate readings at a remote site before transmitting them over a low-bandwidth link.
One test completes this criterion: what the target can express. Rules that can be written in SQL run well there. Decoding a binary format or processing images may require another engine, whatever order is chosen for the rest.
Data that must not enter the target in raw form
Loading first means depositing in the target what the source contains, including, if nothing is filtered, fields that no use requires. For personal data, the GDPR, Regulation (EU) 2016/679, lays down in its Article 5, paragraph 1, point (c), the principle of data minimisation: data that is adequate, relevant and limited to what is necessary in relation to the purposes. Its Article 25, paragraph 2, requires the controller to implement measures ensuring that, by default, only data necessary for each specific purpose is processed; this applies to the amount of data collected, the extent of its processing, the period of its storage and its accessibility.
The text mentions neither ETL nor ELT. We draw a design precaution from it: a raw and complete copy, kept so that processing can be replayed, needs to be justified, with a retention period and access rights. Qualifying each case is for the organisation’s data protection officer and lawyers.
The technical answer is a filter at extraction. Take a temporary employment agency that feeds its warehouse from its payroll software. The source file carries the temporary workers’ social security numbers and bank details, which the activity analyses do not use. These columns are removed, or replaced by a meaningless identifier, before loading; the rest is transformed in the target, and the order becomes mixed.
The classification that designates the fields to be removed falls under data governance, and pseudonymisation and anonymisation techniques are compared in our article on data anonymisation. The same reasoning applies beyond personal data, for a trade secret or a field entrusted by a third party.
Who will write the rules, and how compute will be billed
The third criterion concerns people. In ELT, business rules are SQL queries that the target runs, kept in a code project and covered by tests: this is the use that their documentation, read on the same day, assigns to tools such as dbt or SQLMesh. Analysts who know the business and SQL can then own the transformations themselves. In ETL, the engine has its own language or graphical workbench, and an integration team holds the expertise.
An existing pipeline, maintained by a team that has mastered it, is an asset: the order of the letters is not enough to justify rewriting it.
The fourth criterion is the bill. In ELT, each transformation consumes the target’s compute, whether the offering bills it per query, by execution time or by reserved capacity. A table fully recalculated on each run, when only the day’s rows have changed, weighs on this spending without anyone having decided it.
In ETL, spending concerns the transformation engine and the infrastructure that hosts it. Which of the two orders costs less depends on volumes, processing frequency and the billing model. We suggest attributing the compute consumed to each transformation pipeline before comparing, according to FinOps practices.
Mixed pipelines, and the cases where the debate does not arise
The first mixed set-up is the filter at extraction described above. The second separates destinations: ELT feeds the analytical warehouse, while an upstream engine prepares the flows intended for applications that only accept a specific format.
The third concerns storage. When raw files are deposited in a data lake and transformed by an engine that reads this storage, the boundary between “before” and “after” loading fades. The choice of target, covered in our comparison data warehouse or data lake, then matters more than the acronym.
The debate does not arise in two cases. When data must inform a decision within a time frame that batch loading cannot meet, the transformation is done in the stream, as events arrive, and the questions change: they are covered in our article on real-time data architecture. When sources are few, volumes modest and the target single, both orders are suitable, and the time spent deciding between them would be better spent on testing and documenting the rules.
The order of operations is one of the architecture decisions that our data platform offering records in writing, rejected options included. We suggest noting there, for each flow, the place chosen for the transformation, the reason for the choice and the event that would cause it to be reviewed: change of target, new category of data, drift in compute consumed.