ETL to ELT sounds like a tech refresh. It’s not. It’s a cultural one.
In the old world, Informatica / Talend / DataStage jobs ran overnight, owned by a central data team, and tested manually once a year. In the new world, dbt models version with code, Fivetran loads continuously, and the data team is embedded with the business. The tech change is small. The org change is large.
What goes wrong, in our experience, in 80% of ETL-to-ELT migrations?
Ownership. The central data team owns the pipeline, but the business owns the logic. In the new world, the business analyst writes the dbt model and the data team reviews. That’s a re-org, not a migration.
Testing. Informatica jobs have implicit tests (‘if it ran, it’s correct’). dbt tests are explicit (schema tests, data tests, custom assertions). The new world requires more discipline — which is a feature, not a bug, but it feels like overhead to teams used to the old way.
Finance. Informatica was a single licence line. ELT is per-row, per-warehouse-hour, per-connector. The bill is variable, the cost can surprise, and finance doesn’t trust the FinOps dashboard. Get the FinOps right before you cut over the first model.
So how do we de-risk it?
1. Pilot one model. Pick a single business-critical model (e.g., daily revenue), reproduce it in dbt + Snowflake, run it in parallel with the Informatica job for 4 weeks. Reconcile row-by-row. Sign-off by the finance controller. Then expand.
2. Migrate by domain, not by tooling. Migrate all of Finance’s ETL to dbt in one wave, not all of Informatica. Domain ownership beats tooling completeness.
3. Embed analysts. The analyst writes dbt; the engineer reviews. We pair on the first 5 models, then the analyst takes over. This is the cultural change.
4. FinOps first. Set up the cost dashboard, the resource monitors, the query history. The day the first migration lands, finance can see the cost change in real time.
5. Test like production. CI for dbt on every PR. Schema tests (not_null, unique, relationships). Custom data tests (sum of allocations equals total). Snapshots for SCD-2. Tag-based SLAs.
6. Sunset the old. Once a model is in dbt for 30 days with no defects, archive the Informatica job. Half-migrated estates are the worst.
What does the new stack look like?
– Fivetran / Airbyte for source ingestion (300+ connectors, zero infrastructure).
– Snowflake / Databricks for compute.
– dbt for transformation (versioned, tested, documented).
– Airflow / Prefect / Dagster for orchestration.
– Monte Carlo / Soda / Great Expectations for observability.
– Hightouch / Census / Polytomic for reverse ETL.
The same data, the same answers, 10× faster delivery, lower cost, and a team that owns the logic.