Floemdata
Data & ETL

What is ETL? How it differs from ELT

The three steps of ETL, how it differs from ELT, incremental loading and idempotency — plus the part that decides whether anyone trusts the result.

Floemdata · · 7 min read

ETL stands for Extract, Transform, Load: read data from a source, make it usable, write it to a target.

The three steps

  • Extract: read raw data without loading the source and without embedding business logic in the query.
  • Transform: typing, de-duplication, business rules. Monetary precision must survive this step — no silent float rounding.
  • Load: write to the warehouse. A good load is idempotent: running it twice changes nothing the second time.

ETL or ELT?

ELT reverses the order: raw data lands in the target first and is transformed there. Cloud warehouses favour this because compute is cheap. In practice the difference is where the transformation runs; the rule that the source system must not carry analytical load stays the same.

Why incremental loading is not optional

Re-reading all history every night works on small data and fails on large data. Incremental loading keeps a watermark — how far it got last time — and reads only what is new. Two traps: if the source timestamp has no time component, leave a day of overlap; if rows change retroactively, progress on an increasing key rather than on a date.

Reconciliation decides everything

An ETL pipeline is not successful because data moved. It is successful when the number in the warehouse equals the number in the source system's own report. When they differ, the deviation must be displayed rather than smoothed — trust is lost once and reports are abandoned for good.

← All guides