Scorecard
out of 10What matters most to you?
pick up to 3We'll score ETL and ELT on just your priorities, using the side-by-side below.
Side by side
| Criterion | ETL | ELT |
|---|---|---|
| Where the transformation happens | Before loading, outside the destination | After loading, inside the destination |
| Access to raw data | Only what survives the transformation is kept | Raw source data is kept for re-use (stronger) |
| Flexibility when reporting changes | Transformations are fixed in advance; changes can need engineering help | New models can be built without rebuilding the pipeline (stronger) |
| Control over data before it lands | Suited to complex transformations and quality controls before loading (stronger) | Raw data lands first; cleaning or masking happens afterwards |
| Skills a small team needs | Logic lives in the pipeline tool and is fixed in advance | SQL in the warehouse, often with dbt; someone must still own it (stronger) |
| Cost to get started | Depends on the pipeline tool you choose | Storage billed by use; free tier or sandbox to start; one ELT tool has a free plan (stronger) |
| Unstructured data | Needs pre-processing before it is loaded | Can be loaded as-is, including images and video (stronger) |
The contenders in detail
ETL
- Best for
- Data that must be cleaned or masked before it is stored
- Pricing
- Depends on the pipeline tool you choose
- Hosting
- Your own server or a managed pipeline tool
Pros
- Complex transformations and data quality controls happen before loading
- Only cleaned or filtered data reaches the warehouse
- Suits modest data volumes where an existing ETL tool already works
Cons
- Transformations must be defined in advance, so changes can need engineering help
- Only the transformed data reaches the warehouse, so the raw source is harder to revisit
- Building and changing transformations usually means a pipeline tool or a developer
ELT
- Best for
- Teams using a cloud warehouse with someone who can write SQL
- Pricing
- Warehouse billed by use, plus pipeline tool plans
- Hosting
- Managed cloud warehouse, serverless
Pros
- Raw source data is kept, so a new report does not mean rebuilding the pipeline
- Transformations run as SQL in the warehouse, often with dbt
- Unstructured data such as images and video can be loaded as-is
- The warehouse is serverless, with no infrastructure for you to manage
Cons
- Raw data lands before cleaning, so personal or messy data needs handling in the warehouse
- Someone still has to own the SQL transformations, which is a skills cost for a small team
- Storage is billed by use, so costs rise as you keep more data
If your CRM, invoicing app and website all feed your reports, that data has to move somewhere you can query it. The two acronyms you’ll keep seeing, ETL vs ELT, differ only in one thing: when the reshaping happens.
In ETL (extract, transform, load), you clean and shape data before it lands. In ELT (extract, load, transform), you dump the raw file into a warehouse first and reshape it there. For most small teams on a cloud setup, loading first is the sensible default, with a few named exceptions we’ll get to.
What actually differs between ETL and ELT
The difference is purely about order and location. In ETL you pull data from your sources, reshape it to fit the target system’s needs, then load that cleaned version into the warehouse. Only what survives the transformation gets stored.
ELT flips the last two steps: you extract, then load the raw source straight into a data repository such as a cloud warehouse or data lake, and transform it there once it has landed. You keep the original data available to go back to whenever a new report needs a different cut of it.
Think about post for a second. ETL is sorting envelopes by department in the van before they reach the office, so only what fits your filing system arrives. ELT is dumping every envelope into one big tray and sorting them at your desk later, keeping the unsorted pile if you need to refile. That second option works because a modern warehouse can hold the raw stuff without complaining.
If you want the full background on how ETL actually runs, including what the connectors do, read our explainer on what is an ETL tool. For the wider idea of moving data between systems in general, see what a data pipeline is.
Why loading first became the default
This wasn't always possible. Earlier warehouses were less suited to heavy reshaping after loading, so many teams transformed data outside first. Vendor sources date the change to the years between 2015 and 2020, when cloud warehouses like BigQuery, Redshift and Snowflake emerged with serious processing power built in. As one vendor (Airbyte, who sells an ELT tool, so take the framing with that context) puts it, these warehouses "enabled transformation to be done after loading".
Airbyte's view is that most teams now run their in-warehouse transformations with tools such as dbt, rather than baking everything into the pipeline. You write the reshaping in SQL where your data already sits, test it against real rows, and change it whenever reporting needs shift without touching the load step at all.
Cost to start is low enough that you can trial the pattern before committing. BigQuery, for instance, is serverless (Google runs the infrastructure) and offers a free usage tier or no-cost sandbox so you can begin loading and querying data in minutes, at the time of writing. That makes "just load it and see" a genuinely cheap experiment rather than a big purchase decision.
If you're still new to what a warehouse is and why ELT depends on one, our data warehouse explainer covers that.
How they compare where a small team feels it
Raw access is ELT's quiet win. Because the source data sits in the warehouse untouched, your analyst can build a new model against it without asking anyone to rebuild or reconfigure the pipeline. In ETL, only what you decided to transform and keep survives into the warehouse, so a new report often means going back to the tool that fixed those transformations.
That's the flexibility flip. ELT lets reporting needs change without engineering work; ETL locks in your decisions up front, which can limit how easily you pivot later. For a team whose questions evolve every quarter (and most do), that difference shows up as either a quick afternoon or an invoice from whoever maintains the tool.
ETL does have a real place where it wins: when the data itself is messy at the source and needs complex reshaping, joins across many systems, or strict quality controls before anything is stored. If your sources produce junk that would poison downstream reports if loaded raw, transforming first is defensible.
Skills-wise, ELT pushes the work toward SQL in the warehouse (usually via dbt), which an analyst or a commercially-minded ops lead can own. ETL tends to push logic into the pipeline tool itself, where changes often need someone who set it up. We'll be honest: both approaches still require someone to own the transformations. ELT just makes that ownership less brittle and more accessible.
On cost, storage is cheap but not free. BigQuery bills by use; for 1 TiB of active logical storage in us-central1 you'd pay $23.552 USD a month (a US dollar price for a US region). After 90 days unchanged, data drops to a lower long-term rate. Fivetran, a connector vendor, advertises a free plan to start.
Airbyte's view is that unstructured data (images, video, social feeds) also leans toward ELT: it can be loaded into a lake as-is and queried later, whereas ETL wants to pre-process everything before landing. If your "data" includes anything beyond tidy rows, that matters.
When ETL is still the right call
We'd rather not turn this into an ELT cheerleading exercise, so here's where ETL genuinely wins, in our experience.
If personal or sensitive data must be filtered, masked or reduced before it reaches a shared warehouse, doing that at extract time is cleaner and lower risk than loading raw rows and then scrubbing them inside the warehouse under UK GDPR pressure. That's judgement, not a rule; take proper advice on your specific data, but the instinct to keep certain fields out of a broadly accessible store before they land is sound.
Where transformations are genuinely complex (multi-step joins, validation rules, deduplication logic that must be right every single run), ETL's "do it in the tool where you can test and enforce" model fits better than improvising SQL after the fact.
And if your volumes are modest, your sources are tidy, and you already have an ETL tool working well, there is no obligation to rip it out for a warehouse. The cost of switching usually beats the benefit when nothing is broken.
Which should your business pick
For most small teams with a cloud warehouse (or one they're willing to trial), our default recommendation is ELT: load first, transform in SQL inside the warehouse, and keep the raw data for whatever report comes next week. It's cheaper to start, easier to maintain by non-engineers, and it doesn't lock you into a set of fixed assumptions about what your data means before you've looked at it.
Choose ETL when the data must be controlled before landing: personal information that shouldn't sit raw in a shared store, or messy sources that need heavy pre-cleaning to be usable at all. In those cases, doing it right at extract time is worth the extra rigidity.
The deciding factor is who will own and maintain the transformations over the next two years. If an analyst or ops lead can write SQL and test dbt models against real data, ELT fits your team's reality. If you'd rather hand that to a specialist tool and keep it out of the warehouse, ETL still has its place.
If you want to see a working warehouse of ours, look at our data warehouse case study. And if you're weighing up which approach fits your specific sources and reporting needs, we can walk through it with you: book some time via our data service page and bring the question of what should stay raw.
In practice · Case study
Building a Data Warehouse
Turning five disconnected platforms into one source of truth.
Read the case studyKey terms
- ETL
- Extract, transform, load: the process of copying data from its sources, reshaping it, and then loading it into a database or warehouse.
- ELT
- Extract, load, transform: the same three steps in a different order, with the data loaded raw and reshaped inside the destination.
- Data warehouse
- A database built for reporting and analysis, where data from many sources is stored in one place so it can be queried together.
- Transformation
- Any change made to data on its way to a report or store, such as removing duplicates, renaming fields or joining tables.
- dbt
- A tool for writing and running transformations as SQL inside a warehouse, used by many teams that work in the ELT pattern.
Frequently asked questions
Is ELT replacing ETL?
Do I need a data warehouse to use ELT?
Can a small team do ELT without a data engineer?
What is dbt and where does it fit?
Is raw data in the warehouse a GDPR problem?
Sources
- ETL vs ELT: The Key Differences, Airbyte (accessed 12 Nov 2026)
- BigQuery overview, Google Cloud (accessed 12 Nov 2026)
- BigQuery pricing, Google Cloud (accessed 12 Nov 2026)
- Fivetran pricing, Fivetran (accessed 12 Nov 2026)