ETL vs. ELT – Definition and Differences
Compare transform-then-load with load-then-transform on reprocessing, compute, governance, and why collected data favours one.
The difference is one letter and one property: whether transformation happens before the destination or inside it. That property decides whether history can be reprocessed.
1. The Two Orders
ETL — extract, transform, load. Data is shaped before it reaches the destination, which stores only the result.
ELT — extract, load, transform. Raw data lands first and transformation runs in the destination, which stores both.
2. Key Differences
| Dimension | ETL | ELT |
|---|---|---|
| Where transformation runs | A separate processing tier | The destination |
| Raw data retained | Usually not | Yes, by construction |
| Reprocessing history | Requires re-extraction | Re-run the transformation |
| Compute needed | A dedicated tier | The warehouse's, elastic |
| Storage used | Less | More |
| Governance before landing | Possible | Data lands as-is |
| Time to first load | Slower | Faster |
3. Why ELT Became Common
Warehouse compute became cheap and elastic, and storage cheaper still. Once transformation inside the destination costs less than operating a separate tier, loading raw and transforming in place is simply better.
The reprocessing property then comes free, which turns out to matter more than the cost saving.
4. Why Collected Data Strongly Favours ELT
A web page fetched today cannot be fetched as it was tomorrow. If a transformation discards the original and the parsing is later found wrong, that data is unrecoverable.
Landing raw payloads first makes extraction improvable retroactively. For sources that cannot be re-read, that is not a preference — it is the difference between a fixable mistake and a permanent one.
5. Where ETL Is Still Correct
- Regulatory constraints on what may be stored, requiring reduction or anonymisation before landing.
- Personal data that should be minimised at the boundary rather than warehoused raw.
- Constrained destinations without the compute to transform.
- Very high volume where landing everything is genuinely uneconomic.
The privacy case is the strongest: minimising before storage is a legitimate reason to transform early.
6. The Hybrid Most Teams Run
Land raw with a short retention window, transform into cleaned and modelled layers with longer retention, and apply reduction or anonymisation at the boundary where personal data is involved.
That satisfies reprocessing and minimisation simultaneously, which neither pure pattern does.
7. Both Need Idempotency
Whichever order, runs are repeated after failures and fixes. Deterministic identifiers with upserts, or partitioning by period and replacing a partition wholesale, are what make a re-run routine.
Without it, retrying is a data-quality decision every time.
8. Summary
ETL transforms before loading; ELT loads raw and transforms in the destination, retaining the original.
Collected web data favours ELT because sources cannot be re-read and extraction always improves — with early transformation reserved for privacy minimisation and genuine storage constraints, and idempotency required in either case.