ai-sre-incident-response – skill DevOps & Hạ tầng cho agent AI

AI SRE Incident Response

Apply SRE rigor to AI systems where incidents include quality regressions, unsafe outputs, and budget explosions.

When to Use This Skill

• An LLM endpoint begins returning degraded or hallucinated answers
• Token spend spikes beyond budget thresholds
• A model provider goes down and traffic must fail over
• Safety guardrails fire at abnormal rates
• A new model deployment causes latency or accuracy regression

Prerequisites

• Prometheus and Alertmanager deployed with scrape targets for AI services
• Grafana dashboards for golden signals (latency, error rate, cost, quality)
• On-call rotation configured in PagerDuty, Opsgenie, or equivalent
• Runbook repository accessible to responders
• Rollback mechanism for model and prompt versions (GitOps or feature flags)

AI Incident Classes

• Availability incident: model/provider unavailable, timeout storm.
• Quality incident: answer accuracy or tool success drops below SLO.
• Safety incident: harmful or policy-violating outputs increase.
• Cost incident: unexpected token or provider spend spike.

Severity Framework

| Severity | Criteria | Response Time | Notification |

|----------|----------|---------------|--------------|

| SEV1 | User-facing outage, compliance risk, data leak | 5 min | Page on-call + incident commander |

| SEV2 | Major degradation in key flows | 15 min | Page on-call |

| SEV3 | Limited impact or internal-only issue | 1 hour | Slack alert |

| SEV4 | Cosmetic or low-priority regression | Next business day | Ticket |

Golden Signals for AI Services

• Request success rate
• Latency (queue + generation + tool execution)
• Hallucination/groundedness proxy metrics
• Cost per minute and per tenant
• Guardrail violation rate

Prometheus Alert Rules

The thresholds below are illustrative. Calibrate them to the declared SLO, metric units, tenant budgets and workload; a groundedness proxy is not a verified quality measurement. Counter rates are per second, so the cost-per-minute alert explicitly multiplies by 60.

# prometheus-ai-alerts.yaml
groups:
  - name: ai-service-alerts
    rules:
      - alert: ModelEndpointDown
        expr: up{job="llm-inference"} == 0
        for: 2m
        labels:
          severity: sev1
        annotations:
          summary: "LLM inference endpoint {{ $labels.instance }} is down"
          runbook_url: "https://runbooks.internal/ai/model-outage"

      - alert: HighHallucinationRate
        expr: |
          rate(llm_hallucination_detected_total[10m])
          / rate(llm_requests_total[10m]) > 0.15
        for: 5m
        labels:
          severity: sev2
        annotations:
          summary: "Hallucination rate above 15% for {{ $labels.model }}"
          runbook_url: "https://runbooks.internal/ai/quality-regression"

      - alert: TokenCostExplosion
        expr: |
          60 * sum(rate(llm_token_cost_dollars[5m])) by (tenant)
          > 0.50
        for: 3m
        labels:
          severity: sev2
        annotations:
          summary: "Token spend exceeds $0.50/min for tenant {{ $labels.tenant }}"
          runbook_url: "https://runbooks.internal/ai/cost-spike"

      - alert: LatencyP95Exceeded
        expr: |
          histogram_quantile(0.95,
            rate(llm_request_duration_seconds_bucket[5m])
          ) > 5
        for: 5m
        labels:
          severity: sev2
        annotations:
          summary: "LLM p95 latency exceeds 5s for {{ $labels.service }}"

      - alert: GuardrailViolationSpike
        expr: |
          rate(llm_guardrail_violations_total[10m])
          / rate(llm_requests_total[10m]) > 0.05
        for: 5m
        labels:
          severity: sev1
        annotations:
          summary: "Guardrail violations above 5% for {{ $labels.model }}"
          runbook_url: "https://runbooks.internal/ai/safety-incident"

      - alert: ModelQualityDrop
        expr: |
          llm_eval_score{metric="groundedness"} < 0.70
        for: 10m
        labels:
          severity: sev2
        annotations:
          summary: "Groundedness score dropped below 0.70 for {{ $labels.model }}"

      - alert: ProviderErrorRateHigh
        expr: |
          rate(llm_provider_errors_total[5m])
          / rate(llm_provider_requests_total[5m]) > 0.10
        for: 3m
        labels:
          severity: sev2
        annotations:
          summary: "Provider {{ $labels.provider }} error rate above 10%"

Response Playbooks

Model Outage Runbook

TRIGGER: ModelEndpointDown fires for > 2 minutes
RESPONDER: On-call AI platform engineer

1. Acknowledge alert in PagerDuty.
2. Check provider status page (e.g., status.openai.com).
3. Verify network connectivity:
     curl -s -o /dev/null -w "%{http_code}" https://api.provider.com/health
4. If provider is down:
     a. Enable fallback model route in gateway config.
     b. kubectl set env deployment/llm-gateway FALLBACK_ENABLED=true
     c. Verify fallback traffic is flowing via Grafana dashboard.
5. If self-hosted model is down:
     a. Check pod status: kubectl get pods -l app=llm-inference -n ai
     b. Check GPU health: kubectl logs -l app=llm-inference --tail=50
     c. Restart if OOM: kubectl rollout restart deployment/llm-inference -n ai
6. Request the existing deployment-freeze mechanism for the exact affected scope.
   A Kubernetes annotation by itself does not enforce a freeze; verify that the configured controller/pipeline actually honors it.
7. Communicate ETA in #incident-channel.
8. When resolved, unfreeze and run smoke tests.

Quality Regression Runbook (Hallucination Spike)

TRIGGER: HighHallucinationRate or ModelQualityDrop fires
RESPONDER: On-call AI engineer + ML lead

1. Acknowledge alert. Open incident ticket.
2. Identify scope:
     - Which model version? Check deployment metadata.
     - Which routes/tenants affected? Filter by labels in Grafana.
3. Check recent changes:
     - Model version promotion in last 24h?
     - Prompt template changes in last 24h?
     - Retrieval index rebuild in last 24h?
4. If recent model change:
     kubectl rollout undo deployment/llm-inference -n ai
5. If recent prompt change:
     Prepare a reviewed revert on a topic branch, verify it in staging, and use the protected deployment path; do not push a production rollback blindly.
6. If authorized, increase redacted trace sampling for the affected route only after checking personal-data exposure and the remaining storage/cost budget:
     kubectl set env deployment/llm-gateway TRACE_SAMPLE_RATE=1.0
7. Run the existing version-bound evaluation suite against an authorized isolated copy of the model/prompt configuration using synthetic or redacted cases.
   No run_evals.py helper is bundled; never describe a production call as offline.
8. Confirm metrics return to baseline before closing.

Token Cost Explosion Runbook

TRIGGER: TokenCostExplosion fires
RESPONDER: On-call platform engineer

1. Identify top consumers:
     Query: topk(10, sum(rate(llm_token_cost_dollars[15m])) by (tenant, model, route))
2. Check for runaway loops:
     - Agent retry storms (exponential token growth per request)
     - Missing max_tokens caps on new routes
     - Cache bypass due to config change
3. Apply immediate caps:
     kubectl patch configmap llm-quotas -n ai --patch '
       data:
         max_tokens_per_request: "4096"
         rpm_limit: "60"
     '
4. Review existing cache isolation before any enablement: tenant/user identity, permissions, invalidation and private-content retention must be enforced.
   Do not turn on a generic shared semantic cache during an incident.
5. If authorized, route only to an already approved fallback meeting the declared quality, privacy and cost constraints. Review the exact change before updating gateway configuration.
6. Notify affected tenants of temporary limits.
7. Open postmortem with cost attribution analysis.

Escalation Procedures

Level 1 (0-15 min):  On-call AI platform engineer
Level 2 (15-30 min): AI platform team lead + affected product owner
Level 3 (30-60 min): Engineering director + security (if safety incident)
Level 4 (60+ min):   VP Engineering + legal (if compliance/data incident)

Safety incidents always start at Level 2 minimum.
Provider-side incidents: open support ticket immediately at Level 1.

Detection Queries (PromQL)

# Request success rate by model
1 - (
  sum(rate(llm_requests_total{status="error"}[5m])) by (model)
  / sum(rate(llm_requests_total[5m])) by (model)
)

# Cost per successful answer
sum(rate(llm_token_cost_dollars[5m])) by (route)
/ sum(rate(llm_requests_total{status="success"}[5m])) by (route)

# Hallucination rate trend (1h window, 5m steps)
rate(llm_hallucination_detected_total[1h])
/ rate(llm_requests_total[1h])

# Latency breakdown by stage
histogram_quantile(0.95, rate(llm_retrieval_duration_seconds_bucket[5m]))
histogram_quantile(0.95, rate(llm_generation_duration_secon

… (nội dung đầy đủ có trong gói tải về)