Pattern 1: Metadata-First Semantic Architecture

• Focuses on abstracting and managing metadata at a centralized point.
• Sits between multiple applications and systems, pulling and pushing definitions.
• Most suited for mature organizations due to its complexity.

Pattern 2: Centralized semantic layer

• Gained traction with the rise of data lakehouses, addressing challenges in standardizing business context and meaning across disparate datasets.
• Often used when there’s a need to align metadata across large datasets, as seen in a case study of a global retailer struggling to consolidate store performance data across 40K locations.
• This pattern integrates with enterprise data platforms, leveraging standardized metadata for better data product management and lineage tracking.

Pattern 3: Built-for-purpose Semantic Architecture

• Tailored to specific tools or use cases (e.g., CRM in Salesforce and it has a semantic layer, NetSuite Analytics Data Warehouse has its semantic layer, etc).
• Considered suboptimal due to risks like semantic drift, where definitions across systems may diverge over time.
• More cons than pros, lacking the scalability and consistency of the other two patterns.

Source: Enterprise Knowledge (Data PDX)