Data Lineage for Machine Learning

Lineage helps teams trace where a dataset came from, which jobs used it, and what outputs those jobs produced. OpenLineage models Jobs, Runs, and Datasets, with events describing run state and inputs or outputs. Optional facets can add metadata, including source-code location or more precise job-to-dataset edges. This information is useful for impact analysis, debugging, and reproducibility. If a training dataset changes, lineage may show which model-training run consumed it. If an upstream table has a defect, lineage can help identify downstream tables and jobs. A model registry or experiment tracker may add model and dataset identifiers; a lineage backend then connects the recorded events. There are important limits. Instrumentation must be configured and events must be sent; OpenLineage’s Python source-code-location facet is disabled by default unless enabled. A missing event or metadata facet creates a gap. Events describe reported relationships, not whether source values were accurate, whether a transformation was appropriate, or whether all systems participated. “Complete lineage” is therefore an operational claim that should be tested, not assumed. Point-in-time reproducibility also needs more than a graph. Record immutable dataset versions or snapshots, code revision, parameters, environment, and relevant model artifacts. Retention policies may remove old data even if an event still names it. Use lineage as one evidence layer and verify the referenced versions remain accessible and match the training run.