# SLO + alerting for the inference platform.
#
# SLI:  TTFT p95 over a rolling 5m window.
# SLO:  95% of requests served with TTFT < 1s.
# The `release: kube-prometheus-stack` label makes the Prometheus Operator pick
# this rule up (matches the default ruleSelector).
apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
metadata:
  name: inference-slo
  namespace: inference
  labels:
    release: kube-prometheus-stack
spec:
  groups:
    - name: inference.sli
      interval: 15s
      rules:
        # Recording rules — precompute the SLIs so dashboards and alerts share
        # one definition and stay cheap to evaluate.
        - record: job:vllm_ttft_p95_seconds:5m
          expr: histogram_quantile(0.95, sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le))
        - record: job:vllm_ttft_p50_seconds:5m
          expr: histogram_quantile(0.50, sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le))
        - record: job:vllm_tokens_per_second:1m
          expr: sum(rate(vllm:generation_tokens_total[1m]))

    - name: inference.slo
      rules:
        # Primary SLO alert: TTFT p95 over budget for a sustained window.
        - alert: InferenceTTFTSLOBreach
          expr: job:vllm_ttft_p95_seconds:5m > 1
          for: 5m
          labels:
            severity: warning
            service: mock-vllm
          annotations:
            summary: "Inference TTFT p95 over SLO (>1s)"
            description: "TTFT p95 has been {{ $value | humanizeDuration }} for 5m (SLO: <1s). The autoscaler should be adding replicas; if it isn't, check KEDA and the Prometheus triggers."

        # Saturation early-warning: queue persistently deep means TTFT is about
        # to breach even if it hasn't yet.
        - alert: InferenceQueueBacklog
          expr: sum(vllm:num_requests_waiting) > 5
          for: 2m
          labels:
            severity: warning
            service: mock-vllm
          annotations:
            summary: "Inference request queue backing up"
            description: "{{ $value }} requests waiting for a decode slot for 2m — scale-out signal."