Visual Debugging for Data-Intensive Programming: Evidence-Driven Reasoning

A new arXiv paper by Chen, Zhu, and Faust (submitted to IEEE VIS) tackles the underexplored area of visual debugging in data-intensive programming. Through semi-structured interviews with nine practitioners from diverse domains, the researchers identified three key challenges that current debuggers handle only partially: (1) assembling fragmented evidence scattered across artifacts, (2) detecting mismatches between expected and observed outputs, and (3) tracing state evolution across workflow components like data pipelines.

These challenges translate into three concrete visualization requirements: cross-artifact evidence alignment (linking logs, schemas, and code), expectation-grounded comparison (contrasting actual vs. intended behavior), and traceable state evolution (visualizing data transformations step-by-step). The authors argue that existing debuggers lack these capabilities, leaving data-intensive programmers to rely on manual, ad-hoc methods. This paper provides a foundational design space for building future visual debugging tools tailored to data workflows, aiming to reduce cognitive load and accelerate error detection in ML pipelines, ETL processes, and scientific computing.