Coordinate the baseline with ongoing change
A baseline copy and a change stream must join at a defined position. Without a consistent handoff, records can be missed or applied twice while the source continues changing. Use the source technology's documented snapshot and log mechanisms, and preserve enough position metadata to prove coverage.
Throttle backfills to protect the source and monitor replication lag, queue age, and storage growth. Make transformation logic versioned and rerunnable. If loading into a live projection, separate backfill traffic from current updates or apply ordering rules so an older snapshot row cannot overwrite a more recent event.
| Phase | Required evidence | Decision |
|---|---|---|
| Preparation | Ownership and source inventory | Scope is understood |
| Backfill | Coverage and transformation checks | Baseline is complete |
| Parallel validation | Reconciled business results | New path is credible |
| Cutover | Lag and access checks within agreed limits | Traffic may switch |
| Stabilization | Operational and business review | Old path may retire |
Make cutover criteria explicit
Set measurable criteria for completeness, tolerated lag, reconciliation differences, permission behavior, and recovery readiness. Assign a decision owner and define how the team pauses or aborts. A scheduled cutover time should not override missing evidence that a critical consumer will function correctly.
Shadow reads or parallel reports can reveal semantic differences before switching users. Compare results at the correct grain and include difficult cases such as corrections, nulls, historical classifications, and duplicate identities. Keep the comparison window long enough to include the business processes that matter, such as a settlement cycle or periodic close.
Separate rollback from data repair
Reverting an application deployment does not undo writes already committed to a new store or external system. Define the point after which rollback requires reconciliation or a forward repair. If two systems accept writes during transition, specify conflict resolution and avoid assuming dual writes are atomic.
After cutover, monitor business results, access denials, backlog, and exception rates. Preserve the old path only under a controlled retention and access plan, then retire it when the agreed evidence is complete. Finish with an ownership handover, updated runbooks, and a restoration exercise that reflects the new information path.