Migration guide

    Who’s looking after the metrics while the pipelines move?

    A data warehouse migration is an inventory of sources and pipelines, a map of the metrics people use, a parallel run, and a cutover you can check. Rebuild each metric once, or the reports drift apart afterwards. Platforms like Rig handle mapping, parallel runs and cutover, and leave a governed semantic layer on the warehouse you run.

    By Toby Mather · Published

    Key takeaways

    • The same four steps apply whichever warehouse you pick: inventory, map the metrics, run in parallel, then cut over.
    • Inventory the metrics and the people who read them along with the tables, because the reports are what the business checks.
    • Rebuild each business metric once in a semantic layer on the new warehouse, so every report and AI agent reads the same definition.
    • Agree the parallel run checks before anyone moves data: row counts, business totals and permissions, signed off by the people who own the numbers.
    • Pick the destination against the inventory. Snowflake, Databricks, BigQuery and Fabric each publish their own migration tooling.
    1. 01Inventory

      Sources, pipelines, reports and owners

    2. 02Map metrics

      One definition per business number

    3. 03Parallel run

      Counts, totals and access compared

    4. 04Cutover

      Owners sign off, old jobs retired

    5. 05Destination

      Snowflake, Databricks, BigQuery or Fabric

    The checklist runs the same way for every destination warehouse.

    What are the four steps in a warehouse migration?

    Every warehouse move we see follows the same order, whether the source is SQL Server, Redshift, Teradata, Oracle or Azure Synapse. The destination changes the tooling, and the order stays the same.

    Teams get into trouble when they treat the migration as a data copy. The tables arrive, the pipelines run, and three months later finance and sales report different revenue numbers because each team rebuilt the metric in its own report.

    1. Inventory every source, pipeline, scheduled job, report and downstream consumer, with an owner and a last-used date for each.
    2. Map the metrics people actually use to one definition each, and decide where that definition lives on the new warehouse.
    3. Move the data and pipelines, then run old and new side by side and compare row counts, business totals and permissions.
    4. Cut over one wave at a time once the owners sign off, keep a rollback path for an agreed period, then retire the old jobs.

    What goes into the inventory?

    The inventory decides the budget, and usage data is the most important part of it. A report nobody has opened in a year does not need rebuilding, and a spreadsheet that finance refreshes every Monday probably does, even though nobody listed it.

    Capture the dependencies as well as the objects. A view that four dashboards and an AI agent read is a different job from a view that nothing reads.

    • Sources and connectors, with the refresh schedule each one needs.
    • Pipelines, stored procedures and transformation jobs, with their run order.
    • Tables and views, and which reports, apps and agents read each one.
    • Metrics, with the formula, the owner and every place the metric is calculated today.
    • Users, roles and row-level rules, so access can be tested later.

    Where should the metrics live on the new warehouse?

    Put each metric in one semantic layer on the warehouse, and point reports, apps and agents at it. Snowflake has semantic views, Databricks has metric views in Unity Catalog, and BigQuery and Fabric teams often use a modelling layer such as dbt or a Power BI semantic model. The format matters less than having one definition that every tool reads.

    BigQuery is a common destination for teams already on Google Cloud. Google’s BigQuery Migration Service covers assessment, SQL translation, data transfer and validation from sources such as Teradata, Redshift, Oracle and Snowflake. As with every destination, it moves the SQL and the data, and the metric map is still your job.

    How long should old and new run in parallel?

    Run them in parallel until the agreed checks pass for at least one full business cycle that matters, such as a month-end close or a weekly trading report. The length depends on the scope, so agree the exit checks first and set the dates from them.

    Check row counts per table, then business totals per metric, then permissions per role. Totals catch the errors that row counts miss, such as a currency conversion applied twice.

    The checklist

    Use this list as the exit check for each wave.

    • Every source and pipeline has an owner and a decision: move, rebuild or retire.
    • Every metric in active use has one definition and one owner.
    • Row counts match per table, within an agreed tolerance for late-arriving data.
    • Business totals match per metric for at least one full reporting cycle.
    • Roles and row-level rules have been tested with real users.
    • AI agents and service accounts have their own roles on the new warehouse.
    • A rollback path exists and the business owner has signed off.
    • Old jobs and connectors have a retirement date.

    Which destination guide should I read next?

    Each destination has its own guide with the vendor tooling described from official pages. Read the Snowflake migration guide, the Databricks migration guide, or the Azure Synapse guide if you are choosing between Fabric, Snowflake and Databricks. None of them declares a winner, because the right answer depends on the inventory.

    Where does Rig fit?

    Rig is an AI-native data and systems integrator. An in-house team can run this checklist with the destination vendor’s own tools, and a large systems integrator can staff a multi-year programme. Platforms like Rig handle mapping, parallel runs and cutover, and leave a governed semantic layer on the warehouse you already run.

    In practice that means Rig Migrate runs the inventory and reconciliation at a fixed price, Rig Map holds the metric definitions and lineage after the move, and Gateway gives people and AI agents their own roles and an audit trail. Rig Map and Gateway run on Snowflake, BigQuery or Postgres, or on a warehouse Rig hosts for you. Rig’s published planning allowances are 3 to 5 working days for assessment and 2 to 6 weeks per wave for build and parallel validation.

    Tools and approaches

    Snowflake SnowConvert AI
    Snowflake’s SnowConvert AI translates SQL from sources including Teradata, Oracle, SQL Server, Redshift and Azure Synapse into Snowflake SQL.
    Databricks migration documentation
    Databricks publishes migration guides for data warehouses, ETL pipelines and Spark workloads, and recommends Genie Migrate to assess a source system and convert SQL and ETL code.
    BigQuery Migration Service
    Google Cloud documents assessment, SQL translation, data transfer and validation for moving to BigQuery from other warehouses.
    Microsoft Fabric migration
    Microsoft lists Fabric migration paths by source, including the Migration Assistant for Azure Synapse dedicated SQL pools and SQL Server.

    Common questions

    What is a data warehouse migration checklist?

    It is the list of exit checks for each stage of a warehouse move: an inventory with owners, one definition per metric, matching row counts and business totals during a parallel run, tested permissions, and a signed-off cutover with a rollback path.

    Which warehouse should we pick?

    Pick it against your inventory. Fabric suits teams staying in Microsoft, and Snowflake, Databricks and BigQuery each have different strengths. The Snowflake, Databricks and Azure Synapse guides describe each option from official pages, and a migration readiness check can test the choice against your estate.

    How do we move pipelines without losing the metrics?

    Write the metric map before you move anything, so each business number has one definition and one owner. Then compare business totals per metric during the parallel run, as well as row counts per table.

    When do we cut over to the new warehouse?

    Cut over a wave once its checks have passed for a full reporting cycle and the business owner signs off. Keep the old pipelines available for an agreed rollback period, then retire them.

    Who usually runs a data warehouse migration?

    An in-house team can do this with the vendor’s own tools, a systems integrator can staff the programme, or an AI-native data and systems integrator like Rig can run the mapping and cutover.

    Sources

    Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.

    About the author

    Toby Mather, Co-founder and CEO, Rig. Toby co-founded Rig after seeing how hard it was for business teams and AI to work with complex internal data. He was previously co-founder and CEO of Lingumi and CPO at Novakid.

    Talk to Rig

    Check the migration before you commit to it

    An engineer sends a short written view of the scope, systems and risks, then offers a call to go through it.