A small section of a wooden track is moved aside while remaining connected to the intact main structure.
One reporting workload can be assessed on its own. Its dependencies still need to be part of the decision.

When reports arrive late or every change takes too long, proposing a new data platform is easy. But delays can come from the technology, data model, or the way the team works. Before approving a change, establish the cause and the result you expect. If the cause is still unclear, start with checking the reliability of the company’s data.

In my experience, the scope of change mattered most

I’ve been part of several discussions about rebuilding data platforms. For a small change or one bounded part of the system, it made sense to modify the existing setup. When the data model needed substantial reworking or the changes were broader, it paid off to build a new setup alongside the old one and switch over at an agreed point. In those cases, rebuilding from scratch was usually faster and simpler than making major changes to the original platform.

I’m now wondering whether agents powered by language models could change that tradeoff. With clearly defined procedures and tests at several levels, they might make it faster to validate changes. That could make rebuilding an existing platform while it remains in operation more practical. It’s a possibility I would test on a specific part of the system before relying on it.

Changing technology needs a clear reason

I treat reworking a data model and choosing a different technology as separate decisions. Moving between platforms such as Snowflake, BigQuery, and Databricks needs a strategic reason: why we are moving, what the new platform should make easier, and what it should enable us to do next. I want that benefit spelled out before we start comparing products.

For the first comparison, choose one complete flow of data on which to test that direction. It might run from orders through a sales calculation to a morning dashboard and an export for finance. With a workload defined this way, you can compare how to run it, verify the results, and assign responsibility. A single table is too narrow a starting point if you don't know what writes to it or reads from it. An entire platform can include a lot of work you have no reason to change yet.

Three options, one requirement

Keeping the current solution deserves the same consideration as migrating it. If it serves its purpose, has support, and has someone accountable for its risks, keeping it may be a sound choice. A targeted fix makes sense when you can address a specific cause within the existing flow. Moving the workload becomes an option when an important requirement remains unmet and you can identify a bounded part to test elsewhere.

The following example is entirely hypothetical. Its constraints and timing are chosen to explain the decision. They are not a client experience, migration results, or generally recommended thresholds.

A company needs a sales report for the previous business day by 8:00 a.m. every business day, in its agreed business time zone. The definition of sales is already settled. The problem is that the calculation shares a processing window with other jobs, so it cannot reliably meet the morning deadline under the current conditions. For this example, assume that changing the query or the schedule alone would not remove that constraint.

OptionWhat it would mean in this exampleDecision and reason
Keep itContinue using the current calculation and processing window.Does not meet the morning availability requirement.
Apply a targeted fixChange the query or its schedule within the current solution.Under the stated assumptions, this would not remove the shared constraint. Reconsider this option if the constraint can be removed.
Move this workloadCalculate the report in a separate reporting environment.Test as a candidate because it may allow an independent processing schedule. Switching over is not yet approved.

The decision is to test moving sales reporting. The order system keeps accepting orders and remains the source of record. Other parts of the data platform are outside the scope. Replacing the entire platform would require additional reasons that this example does not provide.

Look beyond the morning dashboard to the quarterly export

The scope of a move depends partly on what the same data produces less often. In this example, finance uses a quarterly export built from the same sales data. It might never appear in a typical week's logs. It still needs to be included in testing and in the decision to retire the old jobs.

Alongside the logs, review export schedules and confirm usage with the people responsible. Google Cloud's migration assessment documentation specifically calls out tables used only once every three or six months. A quiet stretch in the logs does not prove that nobody needs the data anymore.

The workload therefore includes the sales calculation, morning dashboard, and quarterly export. Moving it also requires access permissions, job scheduling, and responsibility for handling failed refreshes. Moving the files and tables alone would not deliver working reporting.

What the new solution has to prove

Before cutover, I want to know that the numbers either match or that we understand why they differ. A changed result can be acceptable when we deliberately change the model or a definition. Material differences need an explanation of where they come from and why they reflect the intended change. The people using those numbers to make decisions also need to accept that explanation.

Matching row counts are a useful check, but they don't tell you whether the report still means the same thing. Two tables can have the same number of rows while handling cancellations or assigning orders to days differently.

In our hypothetical move, the definition of sales stays the same. Both reporting paths therefore receive the same finalized source-data snapshot. The team checks:

  • Keys and duplicates: are both paths counting the same orders, each only once?
  • Sales by day and channel: do both use the same monetary precision and business-day cutoff?
  • Exceptions: are refunds, canceled orders, and late corrections preserved?

The finance lead confirms the included order statuses, the meaning of the results, and the quarterly export format. This checks that the report still supports the same decisions after the move.

Everyday operation also has to pass before the switch: an authorized user can open the report, an unauthorized role cannot, and a failed refresh is visible. The report must not present yesterday's successful refresh as fresh data for today.

For illustration, the team requires five consecutive daily refreshes to finish by 8:00 a.m., plus a successful test using representative quarter-end data. Five runs are a specific condition for this decision, not a universal demonstration of reliability. And a quarter-end test is not an actual quarterly close.

Costs belong in the same decision. Compare all three options for the same workload, data volume, and period. Record recurring operating costs separately from one-time work. If the original platform will keep serving other jobs, moving one report may not reduce its bill. You cannot count that entire bill as savings.

Alongside future operating costs, account for the temporary period when both paths run, duplicate ingestion and storage, data transfer, validation, support, and eventual removal of unneeded components. In this example, the sponsor requires recurring costs after the agreed retirement of old components to be no higher than the documented current operating cost of this reporting workload. The budget for running both paths temporarily needs separate approval.

Until the amounts and conditions are known, the switch stays on hold. Changing environments does not, by itself, demonstrate savings.

Rolling back has to account for new data

The CTO decides on the cutover using the data team’s evidence and the finance lead’s acceptance. If any condition is unmet, the switch does not happen.

Before the change, both paths must process the same completed source snapshot. The team records its identifier and the connection versions. It then redirects only the report’s readers. Orders continue to be written to the original order system.

Switch reporting after validation
Illustrative diagram accompanying the article example.

The old reporting path keeps refreshing from the same sources in case something goes wrong. In our example, both paths run for ten business days. If the team finds a serious error in calculations, permissions, or data freshness, it pauses publication of results.

The CTO approves a rollback only after the data team verifies the old path. It must have processed the latest accepted source snapshot, including new orders and corrections. Only then can readers return to it.

This rollback depends on an important boundary: the new environment is used only for derived reporting. It receives no independent business writes or manual corrections that exist nowhere else. If those changes were introduced, simply switching the connection back could lose them. The team would first need to transfer and reconcile them. AWS similarly distinguishes rollback without data changes from rollback after new writes, when the original copy may already be stale.

The team rehearses the rollback before cutover. If it cannot verify the old path, publication stays paused; an unverified report is not a fallback. Ten business days is the observation period chosen for this example. It does not automatically include a quarterly close or authorize retiring the old reporting path when the period ends.

Retiring the old components has its own conditions

After ten business days, the team reviews operations. The old daily reporting job and view can be removed only after the results have been accepted, rollback has been verified, remaining dependencies have been checked, and the definitions and materials needed for recovery have been preserved. A table stays if another job still reads it.

The old quarterly export path stays in service longer: until the finance lead accepts an actual quarterly run on the new path. The earlier test provided evidence for switching the daily report, but it does not justify removing this dependency. A decision to move sales reporting may therefore lead to several different retirement dates. It sets no retirement date for the entire data warehouse.

Write down a decision you can revisit

The record should preserve why an option was chosen and the conditions under which that choice still holds. Architecture decision records, or ADRs, also use an owner, context, and consequences to document a decision. A concise record for this example might read:

SALES-01 — limited validation approved. Cutover remains pending.

Reason and scope: Test separate sales reporting. Under the stated assumptions, neither keeping the setup nor the proposed fix meets the morning deadline. The order system and other jobs stay where they are.

Responsibility and cutover: The CTO owns the technical decision. The finance lead accepts the meaning of the results and the quarterly export. Switch only after the data, operational, and cost conditions are met and rollback to the current original path has been rehearsed.

Review and retirement: Review operations ten business days after cutover. Retire the old quarterly export path only after an actual quarterly run in the new environment has been accepted.

When to revisit: Approved costs are exceeded, the morning deadline is missed, independent writes are introduced in the new environment, or a dependency prevents planned retirement.

This is a completed decision about the next step. It does not claim that validation or migration has succeeded. The completed example and blank worksheet provide room to record responsibilities, evidence, and conditions in full.

If you already understand the cause of the problem and need to compare options and the scope of change, we can work through them as part of data and AI systems architecture. If you still don't know why the numbers or reports are unreliable, a data audit is a suitable first step.