Metrics Monitoring

Installation

The analytics component is configured with the Prometheus integration. The monitoring for Seldon Deploy is based on the Open Source Analytics package which brings in Prometheus (and Grafana) and is required for metrics collection.

Before installing we should set up a recording rules file. Name this model-usage.rules.yml. The contents of this file are given in the last section.

Create a configmap from the file with

kubectl create configmap -n seldon-system model-usage-rules --from-file=model-usage.rules.yml --dry-run -o yaml | kubectl apply -f -

This can be mounted by setting the below extraConfigmapMounts in an analytics-values.yaml:

grafana:
  resources:
    limits:
      cpu: 200m
      memory: 220Mi
    requests:
      cpu: 50m
      memory: 110Mi

prometheus:
  alertmanager:
    resources:
      limits:
        cpu: 50m
        memory: 64Mi
      requests:
        cpu: 10m
        memory: 32Mi
  nodeExporter:
    service:
      hostPort: 9200
      servicePort: 9200
    resources:
      limits:
        cpu: 200m
        memory: 220Mi
      requests:
        cpu: 50m
        memory: 110Mi
  server:
    retention: "90d"
    extraArgs:
      storage.tsdb.retention.size: 27GB
    persistentVolume:
      enabled: true
      existingClaim: ""
      mountPath: /data
      size: 32Gi
    resources:
      limits:
        cpu: 1
        memory: 2Gi
      requests:
        cpu: 600m
        memory: 1Gi
    extraConfigmapMounts:
      - name: prometheus-config-volume
        mountPath: /etc/prometheus/conf/
        subPath: ""
        configMap: prometheus-server-conf
        readOnly: true
      - name: prometheus-rules-volume
        mountPath: /etc/prometheus-rules
        subPath: ""
        configMap: prometheus-rules
        readOnly: true
      - name: model-usage-rules-volume
        mountPath: /etc/prometheus-rules/model-usage/
        subPath: ""
        configMap: model-usage-rules
        readOnly: true

Other settings in the above are suggested only. Configure to suit your disk availability.

helm repo add seldonio https://storage.googleapis.com/seldon-charts
helm repo update

helm upgrade seldon-core-analytics seldonio/seldon-core-analytics \
    --version 1.4.0 \
    --namespace seldon-system \
    --install
    -f analytics-values.yaml

This Prometheus installation is already configured to scrape metrics from Seldon Deployments. Seldon Core documentation on analytics covers metrics discussion and configuration of Prometheus itself.

It’s possible to leverage further custom parameters provided by the helm charts, such as: * grafana_prom_admin_password - The admin password for grafana to use * persistence.enabled - This provides the configuration to enable prometheus persistence

Bringing your own Prometheus

It is possible to use your own Prometheus instance - see prometheus section in the default values file

seldon-deploy-install/sd-setup/helm-charts/seldon-deploy/values.yaml

If you want to monitor usage of models over time then you need recording rules in place - see below.

Recording Rules

The below is model-usage.rules.yml:

groups:
  - name: model-usage.rules
    interval: 2m
    rules:
      - record: model_containers_average
        expr: label_replace(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) max by (namespace,pod,container,namespace) (avg_over_time(kube_pod_container_info[2m] ))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_memory_usage_bytes
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (rate(container_memory_usage_bytes[2m] )))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_cpu_usage_seconds_total
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (rate(container_cpu_usage_seconds_total[2m] )))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_cpu_requests
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (kube_pod_container_resource_requests_cpu_cores ))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_cpu_limits
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (kube_pod_container_resource_limits_cpu_cores ))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_memory_requests_bytes
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (kube_pod_container_resource_requests_memory_bytes ))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
      - record: model_memory_limits_bytes
        expr: label_replace(sort_desc(sum by (label_seldon_deployment_id,namespace) ((sum_over_time(kube_pod_labels{label_app_kubernetes_io_managed_by=~"seldon-core"}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_seldon_deployment_id) sum by (namespace,pod,container) (kube_pod_container_resource_limits_memory_bytes ))), "name","$1","label_seldon_deployment_id", "(.+)")
        labels:
          type: "SeldonDeployment"
  - name: model-usage-kfserving.rules
    interval: 2m
    rules:
      - record: model_containers_average
        expr: label_replace(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) max by (namespace,pod,container,namespace) (avg_over_time(kube_pod_container_info[2m] ))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_memory_usage_bytes
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (rate(container_memory_usage_bytes[2m] )))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_cpu_usage_seconds_total
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (rate(container_cpu_usage_seconds_total[2m] )))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_cpu_requests
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (kube_pod_container_resource_requests_cpu_cores ))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_cpu_limits
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (kube_pod_container_resource_limits_cpu_cores ))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_memory_requests_bytes
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (kube_pod_container_resource_requests_memory_bytes ))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"
      - record: model_memory_limits_bytes
        expr: label_replace(sort_desc(sum by (label_serving_kubeflow_org_inferenceservice,namespace) ((sum_over_time(kube_pod_labels{label_serving_kubeflow_org_inferenceservice!=""}[2m] ) / scalar(max(sum_over_time(kube_pod_labels[2m] )))) * on(pod,namespace) group_right(label_serving_kubeflow_org_inferenceservice) sum by (namespace,pod,container) (kube_pod_container_resource_limits_memory_bytes ))), "name","$1","label_serving_kubeflow_org_inferenceservice", "(.+)")
        labels:
          type: "InferenceService"

This should be configured with seldon-core-analytics or the same recording rules put into a provided prometheus. Without this you may see warnings about usage.