Both acronyms describe the same three verbs: extract data from a source, transform it, load it somewhere analysts can use it. The only difference is the order of the last two, and that order decides where your computing happens, who can change the logic, and what you can recover when something goes wrong. I will take a side: for most analytics work on a modern cloud warehouse, load first and transform later (ELT) is the better default. Here is the case, and here is where it fails.

What ETL was built for

Classic ETL grew up when warehouses were expensive, fixed-size machines. Storage and compute were bought together in advance, so nobody wanted raw, messy data sitting in the warehouse eating capacity. Instead, a separate server (running tools such as Informatica or SSIS) pulled data out, cleaned and reshaped it, and loaded only the finished tables. This protected the warehouse, and it made sense at the time.

The costs were real though. The transformation logic lived inside a specialised tool, usually owned by a small group of engineers. If an analyst decided that “active customer” should mean 60 days instead of 90, they raised a ticket. And because only transformed data reached the warehouse, the original values were gone. A wrong rule meant going back to the source, if the source still had the data.

What changed

Cloud warehouses such as BigQuery, Snowflake and Redshift separate or elastically scale storage and compute, and storage is cheap. Keeping a raw copy of everything costs little. Meanwhile the warehouse itself is a powerful engine for set-based SQL. So the economics flipped: land the raw data first, in tables that mirror the source, and run transformations inside the warehouse using SQL.

This unlocks several things at once:

  • Re-runnable history. If the definition of net revenue changes, you rebuild from the raw layer instead of re-extracting from a system that may have overwritten old values.
  • Transformations written in SQL, the language analysts already know, and kept in version control. Tools like dbt exist precisely to manage this: models, tests and dependencies as code.
  • Simpler ingestion. Extract-and-load tools can move tables with little custom logic, because they are not deciding what the data should look like.
  • Auditability. Someone questioning a figure can trace it from the report back to the untouched raw row.

Many teams organise the layers as bronze (raw), silver (cleaned) and gold (business-ready), the so-called medallion pattern. It is simply ELT with names on the stages.

Where ETL is still the right call

Here I will argue against my own default, because there are legitimate exceptions.

Sensitive fields. If a source contains Aadhaar numbers, card numbers or health details that analysts should never see, loading them raw into the warehouse spreads the risk. Masking, tokenising or dropping those columns before load is a transform-first step, and under India’s Digital Personal Data Protection Act, 2023, minimising what you store is a sound instinct. You can do ELT for everything else and ETL for these fields.

Heavy non-SQL processing. Parsing images, PDFs or deeply nested logs, or running a machine-learning step, is awkward in SQL. Doing it upstream in code and loading the result is often cleaner.

Legacy or on-premise targets with fixed capacity. If your warehouse really is a fixed box, protecting it with upstream transformation remains sensible.

Data volume reduction at the edge. Filtering out 95 percent of a firehose of sensor readings before sending it over a metered link is a transform-first decision driven by bandwidth, not taste.

What ELT does not fix

ELT can become a swamp. When loading is easy, teams load everything, and nobody cleans up. Costs follow: every transformation runs on metered compute, and a careless full-refresh of a large table every hour can produce a surprising bill. Incremental models, partitioning and a habit of reviewing the most expensive queries each month are part of the deal.

It also does not remove the need for design. Raw tables are not a data model. Someone still has to decide grain, keys and definitions. ELT moves that work to people who can do it in SQL; it does not make the work disappear.

My position, then: start with ELT, keep a raw layer you never edit, and treat the exceptions (privacy, non-SQL processing, hard capacity limits) as deliberate departures with a written reason, rather than defaulting back to the old habit.

Leave a Reply

Trending