Choose placement from access and freshness needs
Federated queries can reduce replication but inherit source availability, query cost, and permission constraints. Replication can support stable analytical workloads but adds synchronization, deletion, and reconciliation responsibilities. Many architectures need both: bounded projections for recurring decisions and governed source access for detailed investigation.
For each dataset, specify the expected freshness, allowed outage behavior, classification, and consumer query pattern. Include the slow path when a source is unavailable. A cached answer should expose its observation time; a missing answer should remain distinct from a legitimate zero value.
| Pattern | Useful when | Review carefully |
|---|---|---|
| Federated access | Current source detail is essential | Source load and access propagation |
| Materialized projection | Repeated low latency reads dominate | Freshness and reconciliation |
| Event subscription | Changes drive decisions | Replay, ordering, and duplication |
Separate context from permission to act
An intelligence layer can assemble the evidence for a decision without receiving unrestricted authority to execute it. Route actions through a service that checks the current resource state, actor permissions, business limits, and required approvals. Evidence available during analysis may be insufficient at execution time.
Record a decision identifier, relevant source versions, policy version, and resulting action status. If an invoice changes after review, reject or revalidate the action instead of treating the old recommendation as permanent authorization. This boundary makes human review and automated workflows easier to reason about together.
Evaluate the design under degraded conditions
Review a source outage, a duplicate event, a revoked user, and a semantic model change before approving the design. For each scenario, state which function stops, which can continue, and how recovery is verified. Cross system transactions usually require explicit coordination or compensating actions rather than assumptions about atomicity.
Use lineage and trace correlation to connect a business result to its inputs and processing history. Define recovery objectives with the business owner, then test restoration and reconciliation against those objectives.