Summary
Key takeaways
- Choose a bounded reporting domain with an accountable owner, a known refresh schedule, and a comparable source of truth.
- Write and sign the business metric contract before porting SQL or changing report pointers.
- Inventory source models, transformations, permissions, dashboards, scheduled exports, and downstream consumers.
- Use a repeatable snapshot or source watermark so legacy and candidate warehouses see the same committed inputs.
- Validate schema, uniqueness, row counts, monetary values, freshness, exports, and allowed and denied access.
- Block cutover when critical daily net revenue differs, even if record counts and other checks pass.
- Estimate bulk backfill separately from change replay, validation, and the cutover rehearsal.
- Give finance, data engineering, BI, security, and release owners explicit approval and stop responsibilities.
- Keep the legacy path current during recovery and observe the next agreed reporting cycle before decommissioning.
When this applies
Use this plan for a bounded reporting domain with a known source of truth, identifiable report consumers, stable ingestion, and accountable finance, data, BI, security, and release owners. The six-week example models a billing report whose scheduled outputs and rollback route can be tested.
When this does not apply
Do not treat the six-week schedule as a fixed commitment for a large historical backfill, a tightly coupled one-time switch, an unavailable business approver, or reporting that requires month-end observation. If the target takes exclusive writes or the legacy source is stale, a simple report-pointer rollback is not sufficient.
Checklist
- Choose one critical report or reporting domain and name its business owner.
- Inventory source tables, transformations, consumers, permissions, and scheduled refreshes.
- Define and approve the metric contract, currency grouping, time zone, and late-credit behavior.
- Choose phased, coordinated, or coexistence cutover criteria with an explicit recovery boundary.
- Configure target service accounts, access controls, and a representative backfill.
- Measure sustained bulk-load throughput and forecast change accumulation and replay time.
- Port versioned transformations and wire the shadow report.
- Align legacy and candidate runs at the same committed source watermark.
- Compare schemas, primary keys, row counts, partitions, and net amounts.
- Test freshness, restricted access, scheduled exports, and dashboard performance.
- Log discrepancies by metric version, watermark, filters, and exact keys.
- Investigate and correct any critical mismatch, then rerun affected periods.
- Rehearse the BI pointer switch, extract refresh, rollback, and incident path.
- Collect named finance, data, BI, security, and release sign-offs before switching.
- Observe the pilot, next scheduled refresh, and agreed business cycle before closing the wave.
Common pitfalls
- Porting SQL before agreeing on a shared business definition of the report.
- Comparing two warehouses that have processed different source watermarks.
- Treating matching row counts or totals as proof of record-level parity.
- Allowing a small finance metric mismatch to pass an exact-tolerance gate.
- Ignoring scheduled exports, restricted-account access tests, or late-arriving records.
- Confusing the backfill duration with the full migration and recovery schedule.
- Assuming a legacy report pointer can restore target-only records.
- Switching without a named release owner, rollback rehearsal, or required sign-offs.
- Decommissioning before the next required reporting cycle and restore checks.
Before moving a critical report to a new data warehouse, define what must match and who can stop the switch. Uvik Software recommends starting with one reporting domain, comparing both warehouses on the same committed inputs, and rehearsing how to restore the legacy report. Finance approves the numbers, data engineering approves freshness, and a release owner makes the cutover decision.
Billing makes a useful first wave because a plausible total can conceal a duplicate credit, and an accurate total can arrive too late for finance. The six-week example and Excel workbook follow both risks from inventory to the first scheduled production refresh. The figures and thresholds are illustrative.
Define the warehouse migration and its first boundary
A data warehouse migration moves the data, transformations, permissions, and reporting workloads that depend on an existing warehouse. The deliverable for the first wave is a business output that can switch to the target and continue meeting its agreed checks. Choose a report or domain with an accountable owner, a known refresh schedule, and a comparable source of truth.
For billing, trace the work from invoices and credit notes through ingestion, transformation, and the semantic layer to dashboards, scheduled exports, and downstream close processes. Find consumers in usage history, service accounts, and BI schedules. Ask finance about monthly spreadsheets and extracts that usage logs can miss. Capture the source and target service accounts, data sensitivity, and fallback for each critical dependency. Google Cloud migration assessment
Write the metric contract before porting SQL. For the example here, daily net invoiced amount sums posted invoice lines after discounts and subtracts posted credit notes, before tax. It groups by posting date and invoice currency. The billing team records the business time zone, the treatment of voids, and how a late credit enters a closed day. Keep the same published definition through cutover. Record any deliberate definition changes in a separate release with its own reconciliation.
Choose where cutover can stop
| Approach | Use it when | Evidence required before switching |
|---|---|---|
| Phased by reporting domain | An owner can switch individual reports or consumers without breaking shared contracts. | A comparable shadow result and a tested route back for each wave. |
| A single coordinated switch | Tightly coupled consumers must move together and the business can pause publication. | A timed rehearsal of all consumers, an agreed outage window, and a validated restore or forward recovery path. |
| Coexistence during a transition | Some workloads must stay on the legacy platform for a business cycle or technical dependency. | A named source of truth for every shared output, a funded dual run, and an exit condition for each dependency. |
Parallel validation can support any of these approaches. Prefer a phased switch for the billing example because the team can change the report pointer for a pilot cohort and keep the existing report current. If the target begins receiving records that the legacy system never sees, confirm where those records can be replayed before treating the old pointer as a recovery route. AWS cutover and rollback guidance
Use a six-week project plan for one domain
The following is a planning example for a bounded billing domain with usable source history, stable ingestion, and an available finance owner. A month-end reporting requirement or a large historical backfill can extend the calendar.
| Week | Work | Required evidence | Accountable owner |
|---|---|---|---|
| 1 | Inventory the source models, report consumers, permissions, and refresh jobs. | Signed metric contract, dependency list, and ranked business-critical outputs. | Finance metric owner |
| 2 | Set up the target data path and backfill a representative date range. | Target service accounts, source checkpoint, backfill timings, and permissions checks. | Data engineering lead |
| 3 | Port transformations and connect a shadow report. | Versioned SQL or dbt changes, schema mapping, and row-level comparisons. | Analytics engineering lead |
| 4 | Run both versions from the same input boundary. | Daily parity log, discrepancy owners, freshness and cost baselines. | Data engineering lead |
| 5 | Rehearse the switch, access checks, and rollback. | Timed practice run, updated dashboard pointers, and written rollback criteria. | Release owner |
| 6 | Switch the approved cohort and observe scheduled runs. | Sign-offs, successful refresh and export checks, and a closeout decision. | Release owner |
Order the work by downstream dependency. Validate a shared customer dimension before comparing revenue by account, and identify the job that refreshes the dimension in each warehouse. Where a stable reporting view or semantic model is available, change its backing source under release control. Test both the query and the scheduled delivery. Exasol data warehouse migration plan
Figure 1. Uvik Software’s illustrative billing migration flow. Validate the legacy and candidate reports against the same source boundary before routing readers to the target.
Put the schedule and backfill in context
Separate the transfer window from the project schedule. A case study may count testing, production rollout, or wider platform work. These three reports give different planning inputs. Record the starting system, the part of the stack that changed, and the exact measurement boundary before drawing an estimate.
| Provider-published case | Reported scale or outcome | What the figure describes |
|---|---|---|
| AWS and Stokrotka | 15 TB in three months | Project kickoff through testing and production. The existing BI front end stayed in place. |
| Snowflake and Flock Safety | About 200 TB in four months, with a four-person project team | The Redshift-to-Snowflake project, with infrastructure, transformations, ingestion, and BI integration work proceeding in parallel. |
| Google Cloud and Vodafone CZ | 46% lower operating costs | A move from fixed-capacity on-premises infrastructure to BigQuery. Compare ongoing operating costs separately from the migration schedule. |
In Uvik Software’s worked backfill model, 2.4 TB at a sustained 200 GB per hour takes 12 hours to land, using decimal units. If changes arrive at 12 GB per hour and replay begins after the load, 144 GB accumulates. A replay process that handles 200 GB per hour while new data keeps arriving clears the backlog at a net 188 GB per hour, or about 46 minutes. Benchmark a representative partition to measure both rates. Before you book the cutover, add time for transformation, validation, reruns, and the rollback rehearsal.
Figure 2. Uvik Software’s worked backfill model. At the modeled rates, 144 GB of accumulated changes remain after a 12-hour bulk load. Replay clears them in about 46 minutes.
Keep the two warehouses comparable during the parallel run
Use a repeatable snapshot or source watermark so that both systems see the same committed inputs. A watermark records how far each pipeline has processed the source, such as a change event offset. The old warehouse and the target may finish their jobs at different times. Validate the last boundary that both reached successfully, then track freshness on each system against its own schedule. If data arrives late or out of order, document the replay window and test that it creates the same result on both sides. Databricks guidance on running warehouses in parallel
Run checks from the lowest level that can explain a discrepancy up to the report a user sees. Start with schemas, primary-key uniqueness, null rates, and row counts by date and key partition. Check amounts by posting date, currency, account, and status, then inspect representative invoice and credit keys. Execute the scheduled export and compare its business fields. Test allowed and denied access with distinct test accounts. Row counts and aggregate totals can each pass while two errors offset one another. A record-level sample and a drill-down by key expose that case. Snowflake data migration and validation · BigQuery migration validation
Save the metric version, watermark, filter, run time, and exact mismatch keys together. First check whether both pipelines reached the same committed source position. Then trace the first differing key through ingestion, transformation, and reporting. Have the downstream owner approve any legitimate rounding or contract change as a scoped exception.
Reconcile a worked revenue example
Use integer cents or a fixed-precision decimal for financial checks. The example requires exact agreement on daily net invoiced amount for each currency after approved rounding. Freshness can have a different tolerance, measured against the published refresh schedule. Here is the first pass for the illustrative billing domain.
| Check | Legacy or target criteria | Candidate observed | Decision |
|---|---|---|---|
| Oct 26 net invoiced amount in USD | $184,905.25 | $184,780.25 | Hold. The candidate is $125.00 lower. |
| Oct 27 net invoiced amount in USD | $192,660.40 | $192,660.40 | Pass. |
| Oct 28 net invoiced amount in EUR | €87,144.18 | €87,144.18 | Pass. |
| Oct 26 invoice key count | 1,248 | 1,248 | Pass. The equal count leaves the amount discrepancy open. |
| Maximum lag on this scheduled validation run | 15 minutes | 9 minutes | Pass. |
| Unauthorized records returned by restricted test account | 0 | 0 | Pass. |
The candidate is $125.00 short, about 0.068% of the legacy total. The agreed tolerance for daily net revenue is zero, so the gap blocks cutover. The example ledger shows a $125 credit note applied twice. After fixing the join, rerun the affected period, compare credit-bearing accounts and currencies, and ask finance to approve the corrected amount and supporting records.
Keep a separate release decision for each gate. A candidate stays on hold if a business-critical amount differs, a restricted account sees unauthorized rows, a required export fails, or any required approver has yet to sign. Set load and performance limits on the same basis. In the sample workbook, a key dashboard query can take at most 2,000 ms at p95 on a matched workload. A team with different user or closing deadlines should set a different limit before running the test.
Find the first mismatch by reporting partition
Stage the two daily extracts in one approved audit database after applying the same business definition and watermark. Require one row per posting day and currency, with amounts in integer cents. This comparison returns missing partitions as well as different amounts. The Oct 26 example returns a delta of minus 12,500 cents.
SELECT COALESCE(s.posting_day, t.posting_day) AS posting_day,
COALESCE(s.currency, t.currency) AS currency,
s.net_cents AS legacy_cents,
t.net_cents AS candidate_cents,
t.net_cents - s.net_cents AS delta_cents
FROM source_daily AS s
FULL OUTER JOIN target_daily AS t
ON s.posting_day = t.posting_day AND s.currency = t.currency
WHERE s.posting_day IS NULL OR t.posting_day IS NULL
OR s.net_cents IS DISTINCT FROM t.net_cents;
Keep duplicate-key and restricted-account checks separate. For a failing partition, compare the underlying keys and posted values before attributing the problem to the warehouse. BigQuery migration validation
Practice cutover and rollback with the same owners
For a reporting-only migration, the first production switch can be a BI connection, view, or semantic model pointer. Keep the legacy warehouse current during the recovery window. Snapshot the existing configuration and save the scheduled job versions. Test cache and extract refresh so that the first dashboard view reflects the target data. If actions on the target create new mutable or enriched records, define how to reconcile or replay those changes before rehearsing rollback. Routing a reader back to the source cannot recover data it never received. AWS guidance on the cutover and rollback stages
| Checkpoint | What happens | Named role |
|---|---|---|
| T minus 60 minutes | Confirm both pipelines reached the cutover watermark. Record the legacy pointer and the restore command or configuration. | Data lead |
| T minus 20 minutes | Review latest parity, role tests, scheduled export, and the state of the legacy refresh. Obtain named sign-offs. | Finance owner, security reviewer, release owner |
| T zero | Switch a small finance cohort. Refresh extracts and run the high-priority dashboard and reconciliation export. | BI owner |
| T plus 30 minutes | Inspect access, freshness, error logs, and response time on the target. Compare sample totals at the confirmed watermark. | BI owner and data lead |
| T plus 90 minutes | Expand only after required checks pass. If a hard gate fails, stop expansion and apply the rehearsed rollback or incident plan. | Release owner |
| Next scheduled run | Confirm the report and export meet their agreed refresh schedule and users have the intended access. | Finance owner and BI owner |
The rollback owner should have explicit authority to stop the cutover. A wrong total or unauthorized access in the pilot cohort triggers an immediate stop. If the legacy system is healthy and caught up, restore the old report pointer, refresh the extracts, and rerun the report and access check. If the legacy system is stale or the target had exclusive writes, pause publication where necessary and use the approved recovery and replay path. Record the period at risk and which reports were affected. A timed rehearsal gives the team better evidence than an estimated recovery duration.
Use the workbook to make the go or hold decision
The accompanying data warehouse migration workbook contains the six-week plan, the populated billing comparison, the cutover and recovery checklist, and named example sign-offs. The example opens on Hold because the Oct 26 amount differs. The row-level checks recalculate when you replace the inputs. Final release readiness also requires an approved recovery checklist and all required sign-offs.
Copy the worksheet for each domain. Add evidence references and assign a person for the finance, data, BI, security, and release decisions. Before decommissioning, agree how long the source must remain available, which records require retention, and which business cycle you will observe. Close the wave only after the next agreed cycle and the archive or restore check have completed. For month-end reports, that observation may extend past the planned switch date.
How Uvik Software connects warehouse and BI responsibilities
Uvik Software’s published data platform case reports that month-end executive reporting went from five days to a four-hour close, measured from reporting timestamps. The same engagement reported 98.7% production pipeline run success from orchestrator history. Those results came from a broader modernization that included Snowflake for warehouse and BI workloads, dbt for transformations and tested metrics, and Databricks for selected heavy-compute and machine-learning workloads. Assign an owner at each boundary so that a failed report can be traced to ingestion, transformation, or the BI definition. Uvik Software’s Snowflake, dbt, and Databricks case study
For platform selection and the cost of running two platforms, see Uvik Software’s data platform modernization guide. Bring one critical report, its current refresh deadline, and the intended switch date when you request a migration readiness review with Uvik Software.