Fine-Grained Provenance for Matching & ETL.

Area of Science:

• Computer Science
• Data Management
• Scientific Computing

Background:

• Existing data provenance tools have limitations in granularity (file-level vs. tuple-level) and scope, failing to adequately support common data science tasks like ETL, record alignment, and matching.
• Current solutions like workflow systems, provenance APIs, and database provenance tools do not effectively trace errors within complex data types (e.g., strings, images) or identify the root causes of discrepancies between code versions or parameter values.

Purpose of the Study:

• To introduce PROVision, a novel provenance-driven troubleshooting tool designed to address the limitations of existing systems.
• To enable scientists to trace errors within data objects during ETL and matching computations, identify sources of errors, and understand the impact of code versions and parameters on analysis results.

Main Methods:

• PROVision extends database-style provenance techniques to capture data equivalences and support optimizations.
• The system enables selective evaluation and traces the extraction of content within data objects, going beyond traditional file- or tuple-level tracking.
• Formalization of extensions, implementation within the PROVision system, and validation through common ETL and matching tasks.

Main Results:

• PROVision effectively supports ETL and matching computations by tracing provenance within data objects.
• The tool demonstrates capabilities in identifying error sources and understanding variations in results due to different code versions or parameter settings.
• Validation confirms the effectiveness and scalability of PROVision for common data science workflows.

Conclusions:

• PROVision offers a significant advancement in data provenance by enabling fine-grained, object-level tracking for troubleshooting.
• The developed system addresses critical needs for scientists in identifying and resolving errors in complex data analysis tasks.
• PROVision's approach enhances the reliability and reproducibility of scientific data analysis by providing deeper insights into the data processing pipeline.