The evolution of performance time series analyzability from application to interpretable evaluation

Abstract

Performance Time Series (PTS) are widely used to monitor system health, yet their analytical value is often limited by a lack of structured event records. This work argues that PTS analyzability is not an inherent property, but a progressive, human-curated and context-dependent property determined by the presence and richness of causal event annotations. We propose a three-level framework – Application Level, Quantifiable Evaluation Level, and Interpretable Evaluation Level – that characterizes the evolution of PTS analyzability from passive data streams into active evaluation infrastructure. By reframing PTS as experimental platforms rather than mere analysis objects, we establish a principled foundation for trustworthy and interpretable algorithm validation.