Comparison
ELTvsETL
ELT
you dumped the source table into the warehouse untouched and did all the shaping afterwards, in SQL, where everyone could read it.
Extract from the source, load it raw, then transform inside the warehouse itself. Cheap elastic warehouse compute is what made this the default: the transform is now SQL in version control that any analyst can read, and the raw landing tables are still there when the definition changes. The bill moves with the work — you are now storing everything and paying to re-read it — which is why ELT shops end up caring about partitioning and incremental models far earlier than ETL shops did.
Full entry →ETL
the data was cleaned and reshaped by a job in the middle, and the warehouse only ever saw the tidy version.
Extract from the source, transform in a processing tier you run, then load the finished shape into the warehouse. It is the older arrangement and it exists because warehouse compute used to be the scarcest, most expensive thing in the building, so you did the work elsewhere. The cost shows up when someone asks a question the transform threw away the answer to: the raw rows are gone, and getting them back means a new pipeline and a wait.
Full entry →