Instrumentation before the next cutover
Dual-write windows fail quietly when telemetry does not cover both old and new paths.
Cutover plans often list feature flags and traffic percentages but forget parity of signals. If only the legacy path emits latency histograms, the new path can degrade without a page until customer support notices.
Define a minimum signal contract before dual-write begins: request rate, error rate, latency percentiles, and a trace sample for the critical span. Apply the contract to both paths with matching attribute names so dashboards compare honestly.
Cardinality budgets matter during migrations. Temporary labels for canary cohorts are useful; unbounded instance IDs are not. Agree on a budget with platform owners before the first canary lands.
Close the dual-write only after a defined quiet period where both paths remain within error budgets. Document the signal proof in the cutover ticket so future migrations reuse the pattern.