release_id: fraud_risk_serving_release_2026_05
endpoint_name: fraud-risk-online
model_name: fraud_risk_xgboost
business_owner: fraud-operations@acme.example
technical_owner: ml-platform@acme.example
runtime:
  serving_platform: kserve
  inference_runtime: triton
  hardware_profile: cpu-standard
  protocol: http
  request_schema_version: fraud-risk-request-v3
  response_schema_version: fraud-risk-response-v3
model_versions:
  current_production: fraud_risk_xgboost:2.4.1
  candidate: fraud_risk_xgboost:2.5.0
feature_service: fraud_risk_service_v3
latency_slo:
  p50_ms: 40
  p95_ms: 150
  p99_ms: 300
batching_policy:
  enabled: true
  preferred_batch_size: 8
  max_queue_delay_microseconds: 500
  applies_to: stateless tabular fraud model
rollout:
  shadow_test:
    enabled: true
    duration: P2D
    traffic_source: mirrored production requests
  canary:
    enabled: true
    initial_traffic_percent: 5
    promotion_steps: [5, 25, 50, 100]
    minimum_observation_window: PT2H
  rollback:
    automatic: true
    owner: ml-platform-oncall@acme.example
    thresholds:
      p95_latency_ms: 180
      error_rate: 0.01
      feature_missing_rate: 0.005
      score_distribution_psi: 0.20
monitoring:
  feature_freshness:
    severity: critical
    metric: maximum_feature_age_minutes
    threshold: 30
  prediction_latency:
    severity: warning
    metric: p95_latency_ms
    threshold: 150
  online_error_rate:
    severity: critical
    metric: http_5xx_rate
    threshold: 0.01
  training_serving_skew:
    severity: critical
    metric: skew_test_failure_rate
    threshold: 0.0
  business_kpi_guardrail:
    severity: critical
    metric: manual_review_rate_change
    threshold: 0.10
approval:
  required_reviewers:
    - fraud-operations@acme.example
    - risk-data-engineering@acme.example
    - ml-platform@acme.example
  promotion_status: pending-review