The Architecture Canontruth · evidence · projection
Pattern

Projected Intelligence Layer

pattern-projected-intelligence-layer · canon/patterns/projected-intelligence-layer/unit.md

Projected Intelligence Layer

Relationship-heavy platforms need graph intelligence without forcing the transactional store to behave like a graph database. Keep the relational DB as truth; project relationship data into a graph asynchronously; reason over the projection. Truth stays singular; traversal and insight stay fast.

Problem

Relationship-heavy products need graph-shaped questions: scene overlap, venue fit, lifecycle importance, promoter relevance, recommendation explanation. The graph is the natural shape for those questions, but it is the wrong place to own transactional truth.

Context

Use this pattern when a product has authoritative business facts in a transactional store and also needs traversal, discovery, ranking, or graph analytics. It fits systems where relationship data has commercial, access-control, ownership, or audit consequences.

Forces

  • Relationship queries want graph traversal.
  • Business facts need one authoritative writer.
  • Intelligence wants fast derived models.
  • Operations need graceful degradation when enrichment lags or pauses.
  • Teams are tempted to fix data where the query is easiest.

Solution

Keep the transactional database as the truth home for relationship facts. Project those facts into a graph asynchronously. Let intelligence read the graph, rank possibilities, and explain relevance. Never let the graph or the recommendation layer become an author of business truth.

Consequences

Core operations keep working when graph enrichment is delayed. Discovery and recommendation can degrade without corrupting ownership, rights, orders, settlements, or approvals. The cost is a more explicit projection pipeline, rebuild path, and authorship discipline.

Failure Modes

  • A graph edge grows a field that does not exist in truth.
  • A support tool fixes relationship state directly in the projection.
  • Recommendations become treated as decisions because they are visible in the UI.
  • Backfills update the graph but not the source relationship record.

Proof Points

recommendation layers are derived.

  • Ampriot — MariaDB owns relationship truth; graph and

Full source pattern: 02-projected-intelligence-layer.md.

Incoming References

Case Study 1