What is ETL?

The three stages

• Extract. Pull data from each source — ad platforms, app analytics, payment systems, CRM — each with its own API, format and rate limits.
• Transform. Reconcile it: currencies, timezones, naming, schema, deduplication. This is where nearly all the work and nearly all the errors live.
• Load. Write the result into a data warehouse for querying.

ETL and ELT

The modern variant reorders the last two: load raw data first, transform inside the warehouse. Cloud warehouses are cheap and powerful enough that this is usually the better arrangement.

The advantage is that the raw data is preserved. If a transformation turns out to be wrong — and it will — you can fix it and reprocess, rather than discovering that the original values were discarded three months ago.

Where pipelines break

Schema changes upstream. A platform adds a field or renames one, and the pipeline either fails loudly or, worse, silently produces nulls.

Timezone handling, which produces persistent offsets that get explained as consumer behaviour rather than recognised as a bug.

Currency conversion applied at inconsistent rates or dates, which quietly corrupts every efficiency metric.

Late-arriving and revised data. Ad platforms restate history; a pipeline that only appends will disagree with the source permanently — see cost aggregation.

Silent failure. The worst category, because a pipeline reporting no error while producing partial data is trusted.