CalejoControl/monitoring/alert_rules.yml

groups:
  - name: calejo_control_adapter
    rules:
      # Application health alerts
      - alert: CalejoApplicationDown
        expr: up{job="calejo-control-adapter"} == 0
        for: 1m
        labels:
          severity: critical
        annotations:
          summary: "Calejo Control Adapter is down"
          description: "The Calejo Control Adapter application has been down for more than 1 minute."

      - alert: CalejoHealthCheckFailing
        expr: calejo_health_check_status == 0
        for: 2m
        labels:
          severity: warning
        annotations:
          summary: "Calejo health check failing"
          description: "One or more health checks have been failing for 2 minutes."

      # Database alerts
      - alert: DatabaseConnectionHigh
        expr: calejo_db_connections_active > 8
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "High database connections"
          description: "Database connections are consistently high ({{ $value }} active connections)."

      - alert: DatabaseQuerySlow
        expr: rate(calejo_db_query_duration_seconds_sum[5m]) / rate(calejo_db_query_duration_seconds_count[5m]) > 1
        for: 2m
        labels:
          severity: warning
        annotations:
          summary: "Slow database queries"
          description: "Average database query time is above 1 second."

      # Safety alerts
      - alert: SafetyViolationDetected
        expr: increase(calejo_safety_violations_total[5m]) > 0
        labels:
          severity: critical
        annotations:
          summary: "Safety violation detected"
          description: "{{ $value }} safety violations detected in the last 5 minutes."

      - alert: EmergencyStopActive
        expr: calejo_emergency_stops_active > 0
        for: 1m
        labels:
          severity: critical
        annotations:
          summary: "Emergency stop active"
          description: "Emergency stop is active for {{ $value }} pump(s)."

      # Performance alerts
      - alert: HighAPIRequestRate
        expr: rate(calejo_rest_api_requests_total[5m]) > 100
        for: 2m
        labels:
          severity: warning
        annotations:
          summary: "High API request rate"
          description: "API request rate is high ({{ $value }} requests/second)."

      - alert: OPCUAConnectionDrop
        expr: calejo_opcua_connections == 0
        for: 3m
        labels:
          severity: warning
        annotations:
          summary: "No OPC UA connections"
          description: "No active OPC UA connections for 3 minutes."

      - alert: ModbusConnectionDrop
        expr: calejo_modbus_connections == 0
        for: 3m
        labels:
          severity: warning
        annotations:
          summary: "No Modbus connections"
          description: "No active Modbus connections for 3 minutes."

      # Resource alerts
      - alert: HighMemoryUsage
        expr: process_resident_memory_bytes{job="calejo-control-adapter"} > 1.5e9
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "High memory usage"
          description: "Application memory usage is high ({{ $value }} bytes)."

      - alert: HighCPUUsage
        expr: rate(process_cpu_seconds_total{job="calejo-control-adapter"}[5m]) * 100 > 80
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "High CPU usage"
          description: "Application CPU usage is high ({{ $value }}%)."

      # Optimization alerts
      - alert: OptimizationRunFailed
        expr: increase(calejo_optimization_runs_total[10m]) == 0
        for: 15m
        labels:
          severity: warning
        annotations:
          summary: "No optimization runs"
          description: "No optimization runs completed in the last 15 minutes."

      - alert: LongOptimizationDuration
        expr: calejo_optimization_duration_seconds > 300
        for: 2m
        labels:
          severity: warning
        annotations:
          summary: "Long optimization duration"
          description: "Optimization runs are taking longer than 5 minutes."