ETL vs ELT: when to prefer each for ML?

WHAT THIS TESTS: Your understanding of data integration patterns and whether you can reason about transformation timing, compute location, and schema rigidity. Interviewers want to see that you know ETL uses a secondary processing server to transform data before loading it into a target database, while ELT loads raw data in its natural state into a data warehouse or data lake and transforms it within the target system. They also want to hear you connect these mechanics to analytical requirements, including the iterative nature of data science work.

A GOOD ANSWER COVERS: First, the process sequence: ETL extracts raw data, transforms it on a secondary processing server, and loads it into the target database only after it meets structural requirements. ELT extracts raw data, loads it directly into the target warehouse or lake, and performs cleansing, enrichment, and restructuring inside that system. Second, the data and schema implications: ETL requires more definition at the beginning, with analytics teams involved from the start to define target data types, structures, and relationships, making it well suited for structured legacy databases. ELT stores unlimited raw data at scale and lets teams analyze it later as required, which supports semi-structured or unstructured sources. Third, the ML relevance: because ELT allows you to interact with and transform raw data as many times as needed, data scientists can experiment with different feature aggregations or filters without rebuilding external pipelines or re-extracting source data. Fourth, the historical context: ETL has been around since the 1970s and was popular with traditional data warehouses, while ELT has become the modern norm with cloud technologies.

COMMON WRONG ANSWERS: Claiming ETL is universally cheaper or faster while ignoring that cloud data warehouses now make ELT efficient at scale. Saying ELT is always superior without acknowledging that legacy systems or strict compliance needs may demand upfront transformation. Describing the difference only as letter order rather than explaining the distinct compute locations and timing. Failing to note that ETL is mainly used for legacy databases while ELT is the current standard.

LIKELY FOLLOW-UPS: How would you handle repeated transformations of the same raw extract for different models? When would you still require a secondary processing server instead of in-warehouse compute? How do you manage schema definition when loading unstructured data into a lake? What governance challenges arise when raw data lands before cleansing?

ONE CONCRETE EXAMPLE: Suppose you are preparing customer transaction records for a fraud detection model. With ETL, you would define the target schema, data types, and cleansing rules on a secondary server before loading, which works well if the source is a structured legacy database with stable columns. If the fraud team later wants to include additional raw fields or change aggregation logic, you must modify the ETL job and reprocess. With ELT, you load the raw transactions into a data warehouse in their natural state, then apply transformations, deduplication, and feature creation inside the warehouse, allowing you to rerun different transformation logic on the same loaded data as the model evolves.

Source: aws.amazon.com