```json { "table_name": "tag_lineage_edges", "primary_key": "id", "candidates": [ { "rank": 0, "partition_key": "dataset_id", "partition_key_label": "E", "partition_key_reasoning": "RPUV=437.75 is in the ideal range (2-100 is preferred, but ~438 is acceptable for 42K rows). Top 3 coverage=29.29% is well below 50%. Semantically, dataset_id represents the entity whose metadata lineage is being tracked — each dataset has its own timeline of tag assignments across tables/columns.", "clustering_key": "first_seen_at", "clustering_key_label": "S", "clustering_key_reasoning": "Temporal ordering axis. Each tag assignment has a first_seen_at timestamp marking when it was first observed. Uniqueness=0.024 (1007 unique values) provides chronological ordering within each dataset's tag history.", "world_line": "dataset_id", "session_column": null, "confidence": "HIGH", "notes": "This is a metadata lineage table tracking when tags (identifier, measure, dimension, etc.) were assigned to columns in various datasets. Partitioning by dataset_id groups all tag assignments for a given dataset together, and sorting by first_seen_at orders them chronologically. The composite entity (dataset_id, table_name, column_name, family, tag) forms the logical grain, but dataset_id is the primary organizing entity." }, { "rank": 1, "partition_key": "consumer", "partition_key_label": "E", "partition_key_reasoning": "RPUV=6003.43 is high but viable for a small cardinality dimension (7 unique values). Top 3 coverage=96.74% indicates skew (card_builder, catalog_service, inspect_columns dominate), but this represents a valid alternative perspective: tracking tag lineage by the system/service that consumed or generated the tags.", "clustering_key": "first_seen_at", "clustering_key_label": "S", "clustering_key_reasoning": "Same temporal ordering as rank=0. Within each consumer's partition, tag assignments are ordered chronologically.", "world_line": "consumer", "session_column": null, "confidence": "MEDIUM", "notes": "Alternative perspective: organize by the consuming service rather than the dataset. Useful for analyzing which services are generating/using which tags over time. The high RPUV and skew reduce confidence, but it remains a semantically valid partitioning for service-centric analysis." } ] } ```