Don't Panic: ETL and ELT

ETL and ELT are two ways of moving data from systems that run the business into a place where it can be analyzed. The letters are less a creed than a map of where the transformation happens. ETL changes data before it reaches the destination. ELT lands source-shaped data first, then transforms it using the destination's compute. The alphabet has caused a great many diagrams to behave as though this were a philosophical dispute. It is a placement choice.

The part worth remembering is the data contract. Before data moves, someone needs to say what identifies a record, what a timestamp means, how updates and deletes arrive, and what freshness is expected. Otherwise a pipeline can copy every available row with impressive punctuality and still construct analytics from a misunderstood source. Machines are highly dependable at repeating an unclear instruction.

Then comes the surprise: an orchestrator can schedule tasks, wait for dependencies, retry failures, and report a cheerful green status without proving the data is correct. Control flow says the machinery ran. Data flow says whether the intended records arrived, passed validation, and became a committed result. Keep those questions separate, because they enjoy being confused with each other.

Incremental work introduces the watermark, a record of how far committed processing reached. It belongs after the destination commit, not after a source read. Timestamps can tie, arrive late, or come from clocks with opinions. A bounded overlap plus deduplication is often more honest than pretending the boundary is perfect. The same caution applies to retries: a repeated run is a replay until the publication method makes it safe.

ETL fits when transformation must happen before landing, such as when restricted fields need removal or the destination cannot do the work. ELT fits when raw replay and destination-side SQL are useful. Both can coexist in one path. A pipeline may protect sensitive values early, then model the remaining records later. Raw data is not an amnesty from access, retention, or quality controls.

Read the Intro for the full route from extraction to recovery. Use Slides when you need the map of data flow and control flow in one view. Keep the Cheatsheet nearby for watermark, idempotency, quality-gate, and backfill decisions. The Practice Reference turns those ideas into a small replay-safe load. The quiz is where the pleasant letters stop being decorative and start asking awkward but useful questions.