AI
New Framework Tackles Root Cause Analysis in Microservices When Data Goes Missing
ARMOR uses self-supervised learning and adaptive fusion to diagnose system failures even when monitoring data streams are incomplete.
Keeping microservice systems running smoothly requires fast diagnosis when things go wrong. Most modern approaches to detecting and fixing failures rely on complete streams of data across multiple sources—metrics, logs, and traces—but in the real world, that data is often incomplete due to network issues or agent failures.
A new research framework called ARMOR, developed by a team including Wenzhuo Qian and colleagues, tackles this common gap by building robust failure diagnosis that works even when crucial monitoring data is missing entirely.
The Problem with Incomplete Data#
Existing systems that analyze microservice failures typically assume all monitoring streams are available. When data does arrive incomplete, they use static placeholder values as a workaround. This approach backfires: filling in the blanks introduces noise that obscures real anomalies and undermines the accuracy of the whole diagnostic process.
How ARMOR Works#
The framework combines two key innovations:
-
Specialized encoders: Instead of treating all data sources equally, ARMOR uses separate pathways for metrics, logs, and traces. This isolation helps the system understand that each data type has its own statistical patterns and isn't immediately comparable to the others.
-
Smart fusion mechanism: Rather than inserting dummy values, ARMOR uses learnable placeholders and dynamic adjustments that prevent incomplete data from throwing off the diagnosis. The approach learns how to compensate for missing information on the fly.
At its core, ARMOR trains using self-supervised learning—it learns patterns from the data itself without requiring extensive labeling. The system then tackles three interconnected tasks: spotting anomalies, classifying what type of failure occurred, and pinpointing which component caused the problem.
Learning with Limited Labels#
The beauty of the self-supervised approach is efficiency: detecting anomalies and localizing root causes require no fault labels at all. The system only needs annotations for the intermediate step of categorizing failure types. This dramatically reduces the annotation burden compared to fully supervised alternatives.
