ETL vs ELT — Why the Industry Switched and What It Means for Your Work
ETL and ELT both move data from source to destination but in a different order. Understanding why the industry shifted from ETL to ELT — and when to use each — is a fundamental data engineering concept.
ETL — Extract, Transform, Load
Traditional ETL extracts data from source, transforms it on a separate compute engine (SSIS, Informatica, custom scripts), then loads the clean, transformed data into the destination.
The transformation happens before loading. The destination only ever sees clean data.
Why it was used: 20 years ago, storage was expensive and databases were slow. Storing raw data was wasteful. Loading dirty data into an expensive Oracle database was unacceptable. So you cleaned it first, then stored only the clean version.
ELT — Extract, Load, Transform
Modern ELT extracts raw data and loads it directly into cloud storage or a data warehouse first. Transformation happens after loading, inside the destination system using its compute power.
On Azure: raw data lands in ADLS Gen2 (Bronze), then Databricks transforms it inside the lake (Silver, Gold). This is ELT — the load happens before the transform.
On AWS: raw data lands in S3, Glue transforms it in-place. Same ELT pattern.
Why ELT won
Cloud storage is extremely cheap — storing raw data costs almost nothing. Cloud compute (Spark, BigQuery, Redshift) is powerful and scalable.
Keeping raw data means you can reprocess with new logic whenever requirements change. With ETL, if you discovered a bug in your transformation after loading, the raw data was gone.
ELT also enables the Medallion Architecture — Bronze is your raw EL layer, Silver and Gold are the T layer. The clear separation of loading and transformation makes debugging and reprocessing straightforward.
When ETL is still correct
ETL still makes sense when: data contains PII that must be masked before storage, transformations are regulatory requirements that cannot be bypassed, or the destination system has strict schema requirements and cannot accept raw data.
For 95% of modern cloud data engineering work: ELT. Keep raw data in Bronze, transform in Silver.