Describe the difference between ETL and ELT and when to choose each

WHAT THIS TESTS: The interviewer wants to know if you understand that ETL and ELT are not just acronyms but architectural choices driven by cost, latency, compliance, and the capabilities of your target system. Specifically, they care whether you recognize who does the work, where the work happens, and what trade-offs that creates for data quality, security, and agility.

A GOOD ANSWER COVERS: Four things in order. First, define ETL as extracting data from sources, transforming it on an external compute layer, and then loading the refined result into the target. Second, define ELT as extracting data, loading it raw into the target system, and then transforming it inside that system using its own compute. Third, explain the decision criteria: choose ETL when you must cleanse, mask, or filter sensitive data before it lands in storage, when data volumes are small enough that external engines are efficient, or when the target database or data lake lacks compute power. Choose ELT when your cloud warehouse is cheaper and faster than external engines, when you want to preserve raw data for future reprocessing, or when schemas evolve quickly and you need flexibility. Fourth, mention real-world economics, such as the shift from traditional ETL tools to ELT in modern cloud warehouses like Snowflake, BigQuery, or Redshift.

COMMON WRONG ANSWERS: Saying ETL is simply old and ELT is simply new without explaining why. Claiming that ELT always replaces ETL, which ignores compliance scenarios where raw sensitive data cannot be loaded. Describing the difference only as the order of letters without discussing compute location or cost. Giving a scenario that does not justify the choice with concrete constraints like data volume, security requirements, or target system limitations.

LIKELY FOLLOW-UPS: The interviewer may ask how you would handle schema drift in an ELT pipeline, how you would secure raw data in an ELT architecture before transformation, or what specific tools you have used for each pattern such as Informatica or Airbyte for ETL and dbt for ELT. They may also ask about cost comparisons between running transformations on Spark clusters versus inside a Snowflake warehouse.

ONE CONCRETE EXAMPLE: Imagine you are building a pipeline for a healthcare company that ingests patient records from multiple clinics. If regulations require that personally identifiable information be masked before it enters the data warehouse, you should choose ETL because the transformation must happen on a secure external compute layer before loading. Conversely, if you are ingesting high-volume clickstream events into a cloud data warehouse and the business wants to experiment with different sessionization models weekly, you should choose ELT because you can land the raw events cheaply and use dbt to iterate on transformations without re-extracting from the source.

Source: Wikipedia: Extract, transform, load