AutoARTS: Taxonomy, Insights and Tools for Root Cause Labelling of Incidents in Microsoft Azure

ARTS - Azure Reliability Tagging System

ARTS (Azure Reliability Tagging System) is a hierarchical taxonomy of root cause tags used to label production incidents of Azure services. Labelling of postmortems enables aggregating root cause categories to identify problem areas, trends, patterns, and risks that may lead to future incidents.

The ARTS taxonomy is comprehensive---it includes many different factors contributing towards thousands of production incidents in Azure, and yet compact enough---it contains only the root causes observed in real incidents. The taxonomy has been developed over last few years and is still growing (slowly) as new root causes are discovered.

Description of the ARTS Taxonomy:

• Visualization: Each node in the hierarchical taxonomy shows a root cause category and it can be clicked to expand its sub-categories (if any). Moreover, hovering the mouse over a node shows a short description of the category.
• CSV File: Each line contains two fields. The "TagName" field is the concise tag for a category, and the "TagDef" field briefly describes the category. Each category is hierarhical, with each level of the hierarchy seperated by a dot. For example, the tag "Architecture.ColdStart" indicates the "ColdStart" sub-category of the "Architecture" category. 

Automatic labelling with ARTS

One can manually label a postmortem with relevant ARTS tags. However, such manual labelling does not scale well and can be inconsistence/error-prone (see paper).

To address this, we developed automated ML-based techniques to analyze the content of a postmortem called AutoARTS in order to (1) recommend ARTS tags relevant to the postmortem, and (2) a small snippet of texts (from the long postmortem) that capture the rootcauses explaining the recommended tags. The work is published at USENIX ATC'23, email us at autoarts-rca-taxonomy at outlook dot com if you are interested in the details.