AWS DMS

CDC latency averages 15-30 seconds in practice, with occasional spikes to 2+ minutes during high transaction volumes. Cold start for new replication tasks takes 5-15 minutes. While acceptable for batch analytics, this violates the sub-2-second agent response requirement when combined with downstream query processing. Multi-AZ deployments help but don't eliminate the inherent batch-processing delays.

DMS uses proprietary transformation rules syntax that requires specialized DBA knowledge. No semantic understanding of business entities — treats customer_id and cust_id as unrelated fields. Schema conversion tools exist but require manual mapping for complex transformations. Teams typically need 2-3 weeks to become productive, and migration scripts are not transferable to other platforms.

Strong IAM integration with fine-grained resource-level permissions and VPC endpoint support for network isolation. Supports column-level filtering during replication. However, lacks native ABAC — relies on IAM roles which cannot evaluate dynamic business context like time-based or data-sensitivity rules. SOC 2 Type II, ISO 27001, and HIPAA eligible with BAA.

Hard lock-in to AWS ecosystem — cannot replicate to non-AWS targets without complete reconfiguration. Migration path to alternatives requires rebuilding entire replication infrastructure. No drift detection for schema changes; replication tasks fail silently when source schema evolves. Limited plugin ecosystem compared to Kafka Connect or Airbyte's connector marketplace.

Excellent source system support (Oracle, SQL Server, MySQL, PostgreSQL, MongoDB) but no native metadata preservation. Data lineage stops at the replication task level — no field-level tracking or semantic relationship preservation. Integration with AWS Glue helps but requires additional configuration. Cannot maintain cross-system entity relationships during migration.

CloudWatch provides basic metrics (latency, throughput, errors) but no query-level cost attribution or semantic impact analysis. Cannot trace which source system changes affected which downstream agent decisions. Audit logs show replication events but not business impact. No built-in data quality monitoring — silent data corruption can persist undetected.

Strong integration with AWS CloudTrail for audit logging and AWS Config for compliance monitoring. Supports data sovereignty through region selection and encryption in transit/at rest. However, no automated policy enforcement for data classification or business rule validation during replication — relies on external governance tools.

CloudWatch integration provides infrastructure metrics but no AI/ML-specific observability. Cannot track semantic data quality or downstream model performance impact. Third-party tools like DataDog can enhance monitoring but require additional integration work. No native support for LLM observability metrics.

99.9% uptime SLA with multi-AZ deployment options. RTO typically 15-30 minutes for failover, which exceeds the 1-hour cap but acceptable for many use cases. RPO near-zero for most replication scenarios. However, no active-active configurations — failover is manual and requires DNS updates for applications.

Basic schema mapping capabilities but no semantic layer integration. Cannot preserve business glossary terms or entity relationships during migration. AWS Glue Data Catalog integration helps but requires manual configuration. No support for ontology standards or automated metadata enrichment.

Launched 2016, mature service with thousands of enterprise customers including major banks and healthcare systems. Proven track record for high-volume migrations (petabyte-scale). Conservative change management with extensive backward compatibility. AWS's reliability track record and enterprise support provide additional stability assurance.