AI-Assisted Root Cause Analysis for Operations

Problem Statement

Operating a 5G Core on Kubernetes generates a huge amount of observability data spread across metrics, logs and traces stores. When something goes wrong, engineers have to manually correlate signals across these different stores to find the root cause, which is slow and requires deep expertise.

Description

Explore how to "connect" LLMs to metrics, logs and traces stores to perform distributed LLM-based debugging: let the LLM query the different observability backends, correlate the signals and propose a root cause.

The idea is also to bring in the Alarms view, so an engineer could tell the LLM "I have an alarm with ID=XYZ" and get back the likely root cause, the correlation with other alerts, and possible remediation steps.

Aspect Description
Signal access Connect the LLM to metrics, logs and traces stores
Correlation Correlate signals across stores to spot issues
Root cause analysis Explain what went wrong from an alarm ID or symptom
Remediation Suggest possible remediation steps
OTel comparison (bonus) Compare how well the LLM performs with OTel-compliant signals vs non-OTel ones

Project Details

Leaders:

  • Swisscom (Joel Studler, Alessio Diamanti, Fabian Schulz)

Hackathon Objectives

  • Connect an LLM to metrics, logs and traces stores for distributed debugging
  • Perform root cause analysis and correlation from an alarm ID
  • Suggest remediation steps
  • Bonus: compare LLM performance on OTel-compliant vs non-OTel signals