ETL, ELT and reverse ETL, in plain English.

Extract, transform, load. Extract is pulling data out of a platform. Transform is reshaping it into something consistent. Load is putting it where it will be queried.

ETL does the transform before the load, so what arrives in storage is already clean. ELT loads the raw data first and transforms it afterwards, inside the warehouse. The same three jobs in a different order, and the order changes what you can do later.

Transforming first means deciding, at the moment you build the pipeline, which fields matter. Anything you did not keep is gone. That is a reasonable trade when the source is a database you control and the questions are known.

Marketing platforms are neither. They add and rename fields without asking, they restate figures for days that have already closed, and the question a client will ask in six months has not been asked yet. Loading the raw data first means a new question can be answered from history you already hold, instead of from the day you thought to start collecting it.

Storage being cheap is the other half of it. At some point it stopped being sensible to throw data away to save space.

It is not free, and the trade is worth stating.

• You store more

  Raw and modelled data both live in the warehouse. Cheap is not the same as free, and a raw layer nobody prunes grows for years.

• The transform becomes the thing to trust

  With ELT the numbers people read are produced by transform logic rather than by the pipeline. If that logic is not readable, nobody can check how a figure was built, and an unreadable transform is worse than none.

• You need somewhere to run it

  The transform has to execute on a schedule, fail loudly and be safe to re-run. That is a job, and it is the one most often underestimated when an agency decides to build this itself.

Reverse ETL is sending modelled data back out to the tools a client's team works in, so a segment built in the warehouse can be used in an email platform or an ad account.

It is worth having once the warehouse holds something the source tools do not, such as what a customer is worth across every channel. It is worth skipping while the only consumer is a dashboard, because pushing data back out before anyone trusts it just spreads a number nobody believes.

If you control the source, the schema is stable and the questions are fixed, transforming first is simpler and there is no reason to store raw data you will never read.

If the sources are platforms you do not control, transform second. Almost all marketing reporting is the second case, which is why this stopped being a real debate for marketing data some years ago. The live question is not which order. It is who maintains the transform.

SightShark loads every client's sources first and transforms them inside a warehouse you own, so the raw history stays yours and the modelling runs where you can read it.

The transform is SQL in your own project rather than logic inside a product you rent, which means you can see how any number was built and it does not leave when we do.