CDC Pipeline Correctness Auditor

Act as a staff data engineer reviewing a business-critical CDC pipeline. Do not assume exactly-once delivery. Identify how the system can silently lose, duplicate, reorder, or resurrect records. Source database: [PostgreSQL / MySQL / SQL Server / other] CDC technology: [Debezium / DMS / Datastream / Fivetran / custom] Transport and sink: [Kafka → Flink → Iceberg, etc.] Table keys and mutation patterns: [primary keys, updates, deletes] Scale: [changes/sec, peak multiplier, table size] Recovery guarantees and RPO/RTO: [requirements] Current design or configuration: [Paste architecture, connector settings, and relevant schemas] Produce: 1. End-to-end correctness model from source commit to queryable sink 2. Failure-mode analysis covering snapshots, log gaps, duplicates, out-of-order events, deletes, schema changes, and connector restarts 3. Idempotency and deduplication design, including stable event identity 4. Source-offset, checkpoint, and replay strategy 5. Reconciliation queries that compare source and sink without full scans where possible 6. Recovery procedure for each failure class 7. Required metrics, alerts, and data-loss indicators 8. A prioritized remediation plan labeled P0–P2 State assumptions explicitly. Separate guarantees provided by each technology from guarantees the full pipeline actually achieves.