ETL to ELT Migration Checklist
A checklist for migrating ETL pipelines to an ELT model that loads raw data first and transforms inside the warehouse. It covers layered design, in-warehouse transforms, reconciliation, cost modeling, and parallel running. Use it to move transformation into the cloud warehouse safely.
When to Use This Checklist
Use this checklist when moving from ETL to ELT. In ETL, data is transformed before loading; in ELT, raw data is loaded first and transformed inside a powerful cloud warehouse. ELT suits cloud platforms like Snowflake, BigQuery, and Databricks, where compute is elastic and keeping raw data is cheap. The shift changes where transformation runs and how pipelines are tested and governed.
How to Use This Checklist
Start with a full inventory of existing ETL jobs so no transformation logic is lost. Design the layered model: raw, staging, and curated. Reimplement transforms as in-warehouse SQL, ideally in a tool like dbt that supports testing and lineage. Treat reconciliation against the legacy pipeline as a hard gate, and run both pipelines in parallel until the new outputs earn trust. The required items protect correctness and auditability; the optional items add lineage and cleanup.
What Good Looks Like
A good ELT migration keeps raw source data intact so transforms can be replayed and audited. Transformation logic lives in version-controlled SQL with automated tests, not in opaque jobs. Outputs reconcile to the legacy pipeline before cutover, and the two run side by side until confidence is high. Compute cost is modeled in advance, lineage spans every layer, and the legacy infrastructure is only retired after a stable soak. Analysts gain faster iteration and full traceability.
Common Pitfalls
The biggest risk is silently changing business logic during reimplementation, so outputs drift from the trusted ETL results; reconciliation catches this. Teams also underestimate warehouse compute cost, since transforms now run on billable warehouse time. Discarding raw data removes the ability to replay or audit. Skipping tests on transformed models lets regressions reach reports. Finally, cutting over before a parallel-run period removes the safety net too early.
Related Resources
Frame the decision with elt-vs-etl-best-practices and structure layers using medallion-architecture. Write transforms to the dbt-style-guide, and protect outputs with data-lineage and data-quality-management.