### 1.2 Key Data Assets | Table | Role | Rows | % of Total | Size | |-------|------|-----:|----------:|-----:| | columns_catalog | dimension | 1,533,257 | 74.6% | — | | job_logs | path | 303,822 | 14.8% | — | | llm_calls | path | 49,360 | 2.4% | — | | column_tags | dimension | 43,853 | 2.1% | — | | tag_lineage_edges | path | 42,024 | 2.0% | — | The **columns_catalog** table dominates the dataset with three-quarters of all records, serving as the central repository of column-level metadata across profiled data sources. The **job_logs** and **llm_calls** tables capture execution telemetry and AI-assisted analysis activity, representing the operational heartbeat of the platform's processing pipeline. Together, these five tables account for 96% of the dataset's volume, with the remaining 15 tables providing classification taxonomies, dataset registries, and lineage tracking infrastructure. ### 1.1 Business Data Landscape This dataset represents the internal telemetry and metadata catalog of a data intelligence platform that profiles external datasets, classifies their structure, and tracks analysis workflows. Core entities include **datasets** (the external data sources being analyzed), **columns_catalog** and **tables_catalog** (the profiled schema elements), **analysis_runs** and **batches** (processing orchestration), and **llm_calls** (AI-powered classification and tagging). The architecture separates catalog metadata (what was profiled) from execution telemetry (how and when profiling occurred), with dimension tables like **column_tags** and **table_classifications** providing the semantic layer that translates raw schema into business-meaningful categories. Path tables such as **job_logs** and **tag_lineage_edges** trace the operational history and dependency chains that connect profiling runs to their outputs. ### 1.3 Data Integration Posture Despite containing 74 detected relationships across 20 tables, zero of these connections are currently join-viable, indicating a dataset in active development where referential constraints have not yet been enforced or where data has been extracted for analysis without preserving foreign key integrity. This fragmentation limits cross-functional analytics: correlating LLM usage costs (**analysis_run_llm_usage**) with specific column classifications (**column_tags**) or tracing job failures (**job_logs**) back to dataset characteristics (**datasets**) requires manual key matching rather than reliable joins. For business leaders, this means current reporting likely operates in vertical slices—analyzing job performance *or* catalog coverage *or* AI costs—rather than synthesizing these dimensions into unified operational dashboards that answer questions like "Which dataset types consume the most LLM tokens?" or "Do classification overrides correlate with job retry rates?"