Jev-Driven SRE Diagnosis: What Worked and What Failed
4 hours ago
- A Jev-driven diagnosis pipeline was developed that collects cluster evidence programmatically and feeds it to Jev for root cause analysis without an LLM agent.
- Across 21 SREGym-Lite faults, the pipeline achieved 80/105 (76.2%) diagnosis passes with a median time of 14.6 seconds.
- The collector organizes Kubernetes objects, events, logs, and resource usage into summaries for Jev, which then selects likely root causes and supporting evidence.
- Jev answers multiple-choice questions based on provided evidence and options, not generating commands or writing reports.
- Two main failure modes were identified: Jev choosing the wrong clue and missing decisive evidence in the pipeline.
- The pipeline's performance is close to GPT-5.6 Sol (medium) but about 7x faster and 200x cheaper.
- Future work aims to handle multi-service and evolving faults by integrating request-level signals and specialized models.