ai-sre-incident-response

AI SRE incident response practices define incident classes, severity frameworks, and golden signals to address LLM outages and quality degradation. They provide specific response playbooks and postmortem requirements to mitigate availability, quality, safety, and cost incidents.

Golden signals for monitoring AI service health identify key metrics across availability, quality, safety, and cost dimensions. These metrics help detect model outages, quality regressions, safety issues, and cost overruns specific to LLM services.

AI incident severity is classified using a clear SEV1-SEV3 severity framework tailored for AI services. This framework categorizes incidents across availability, quality, safety, and cost dimensions to prioritize response actions for degraded model performance.

Yes, SRE response playbooks outline specific actions for cost incidents like runaway LLM API costs. They define mitigation strategies and monitoring metrics to detect and resolve cost overruns within AI services.

AI incident postmortems detail essential elements for thorough incident reviews following safety regressions. They require documenting specific mitigation strategies, monitoring metrics, and response actions taken during the safety incident.

You need a dedicated AI SRE framework when addressing unique AI incident classes like model outages, safety regressions, and degraded quality that traditional DevOps practices do not cover. It establishes golden signals and playbooks specific to LLM services.