Skip to content

Latest commit

 

History

History
192 lines (147 loc) · 11.1 KB

monitoring.md

File metadata and controls

192 lines (147 loc) · 11.1 KB
Raw

ModelMesh Metrics

Overview

Serving runtime pods expose the endpoint /metrics on port 2112 and scheme https and the metrics published by each pod are mostly disjoint, for aggregation by a monitoring framework (e.g. Grafana), except for service-wide metrics with the scope Deployment. These are published by only one of the pods at a given time.

Endpoints associated with a ModelMesh Serving service (modelmesh-serving by default) should be used to track the serving runtime pods' IPs from which Prometheus metrics should be scraped.

To override the default metrics port and scheme, you can set the variables metrics.port and metrics.scheme in the model-serving-config configmap).

Metrics

ModelMesh Serving publishes a variety of metrics related to model request rates and timings, model loading/unloading rates, times and sizes, internal queuing delays, capacity/usage, cache state/LRU, etc. Each serving runtime pod exposes its own metrics that should be aggregated, except for service-wide metrics with the scope "Deployment" that are published by only one of the serving runtime pods at a given time.

Here is the list of metrics exposed by ModelMesh Serving:

Name Type Scope Description
modelmesh_invoke_model_milliseconds Timing Internal model server inference request time
modelmesh_api_request_milliseconds Timing External inference request time
modelmesh_request_size_bytes Size Inference request payload size
modelmesh_response_size_bytes Size Inference response payload size
modelmesh_cache_miss_milliseconds Timing Cache miss delay
modelmesh_loadmodel_milliseconds Timing Time taken to load model
modelmesh_loadmodel_failure Count Model load failures
modelmesh_unloadmodel_milliseconds Timing Time taken to unload model
modelmesh_unloadmodel_failure Count Unload model failures (not counting multiple attempts for same copy)
modelmesh_unloadmodel_attempt_failure Count Unload model attempt failures
modelmesh_req_queue_delay_milliseconds Timing Time spent in inference request queue
modelmesh_loading_queue_delay_milliseconds Timing Time spent in model loading queue
modelmesh_model_sizing_milliseconds Timing Time taken to perform model sizing
modelmesh_age_at_eviction_milliseconds Age Time since model was last used when evicted
modelmesh_loaded_model_size_bytes Size Reported size of loaded model
modelmesh_models_loaded_total Gauge Deployment Total number of models with at least one loaded copy
modelmesh_models_with_failure_total Gauge Deployment Total number of models with one or more recent load failures
modelmesh_models_managed_total Gauge Deployment Total number of models managed
modelmesh_instance_lru_seconds Gauge Pod Last used time of least recently used model in pod (in secs since epoch)
modelmesh_instance_lru_age_seconds Gauge Pod Last used age of least recently used model in pod (secs ago)
modelmesh_instance_capacity_bytes Gauge Pod Effective model capacity of pod excluding unload buffer
modelmesh_instance_used_bytes Gauge Pod Amount of capacity currently in use by loaded models
modelmesh_instance_used_bps Gauge Pod Model capacity utilization in basis points (100ths of percent)
modelmesh_instance_models_total Gauge Pod Number of model copies loaded in pod

Note: The request metrics include labels method and code with the method name and gRPC response code respectively. The code for successful requests is OK.

The best way to visualize the metrics is to use Prometheus to collect them from targets by scraping the metrics HTTP endpoints coupled with a Grafana dashboard. Setup instructions are provided below and involve the following steps:

  1. Set up Prometheus Operator
  2. Create the ServiceMonitor CRD
  3. Import Grafana Dashboard

Monitoring Setup

Set up Prometheus Operator

The Prometheus Operator is the easiest way to set up both Prometheus and Grafana natively in a Kubernetes cluster. You can clone the kube-prometheus project and follow the quickstart instructions to set it up. By default, the operator sets RBAC rules to enable monitoring for the default, monitoring, and kube-system namespaces to collect Kubernetes and node metrics.

Monitor Additional Namespaces

To monitor the modelmesh-serving namespace, in the cloned kube-prometheus repository, add the following to manifests/prometheus-roleBindingSpecificNamespaces.yaml:

- apiVersion: rbac.authorization.k8s.io/v1
  kind: RoleBinding
  metadata:
    labels:
      app.kubernetes.io/component: prometheus
      app.kubernetes.io/name: prometheus
      app.kubernetes.io/part-of: kube-prometheus
      app.kubernetes.io/version: 2.30.2
    name: prometheus-k8s
    namespace: modelmesh-serving
  roleRef:
    apiGroup: rbac.authorization.k8s.io
    kind: Role
    name: prometheus-k8s
  subjects:
    - kind: ServiceAccount
      name: prometheus-k8s
      namespace: monitoring

and to manifests/prometheus-roleSpecificNamespaces.yaml:

- apiVersion: rbac.authorization.k8s.io/v1
  kind: Role
  metadata:
    labels:
      app.kubernetes.io/component: prometheus
      app.kubernetes.io/name: prometheus
      app.kubernetes.io/part-of: kube-prometheus
      app.kubernetes.io/version: 2.30.2
    name: prometheus-k8s
    namespace: modelmesh-serving
  rules:
    - apiGroups:
        - ""
      resources:
        - services
        - endpoints
        - pods
      verbs:
        - get
        - list
        - watch
    - apiGroups:
        - extensions
      resources:
        - ingresses
      verbs:
        - get
        - list
        - watch
    - apiGroups:
        - networking.k8s.io
      resources:
        - ingresses
      verbs:
        - get
        - list
        - watch

Increase Retention Period

By default, Prometheus only keeps a 24-hour history record. To increase the retention period, modify manifests/prometheus-prometheus.yaml by adding:

spec:
  ...
  resources:
    requests:
      memory: 400Mi
  # To change the retention period to 7 days, add the line below
  retention: 7d
  ...

Other configurable Prometheus specification fields are listed here.

Create the ServiceMonitor Resource

ServiceMonitor is a custom resource definition provided by Prometheus Operator and is leveraged by ModelMesh for monitoring pods through the modelmesh-serving service.

Create a ServiceMonitor to monitor the modelmesh-serving service using the definition found here.

kubectl apply -f servicemonitor.yaml

After the ServiceMonitor is created, the Prometheus operator will dynamically discover the pods with the label modelmesh-service: modelmesh-serving and scrape the metrics endpoint exposed by those pods.

Note: By default, when the ModelMesh controller is started, the ServiceMonitor is checked. If it exists and metrics.enabled is true, a ServiceMonitor resource will be created for monitoring ServingRuntime pods.

If you have an alternative solution to collect the metrics, you can disable the creation of ServiceMonitor by setting the configuration metrics.disablePrometheusOperatorSupport to true in the model-serving-config configmap.

Import the Grafana Dashboard

To access Grafana and visualize the Prometheus-monitored data, follow the instructions here and import the pre-built dashboard we provide using this guide.

Troubleshooting

If the ModelMesh Serving metric(s) are missing in the monitoring UIs:

  • Check if the Serving Runtime pod is up and running.

    • Check if the annotations are configured in the Serving Runtime deployment:

      kubectl describe deployment modelmesh-serving-mlserver-1.x -n $NAMESPACE | grep "prometheus.io"

  Annotations:  prometheus.io/path: /metrics
                prometheus.io/port: 2112
                prometheus.io/scheme: https
                prometheus.io/scrape: true

      Note: The configured metrics.port must be listed in the annotation prometheus.io/port. The default port is 2112.

  • Check if the configured metrics.port is exposed through the service modelmesh-serving:

      kubectl get svc modelmesh-serving -n $NAMESPACE

  NAME               TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)             AGE
  modelmesh-serving  ClusterIP   ..........      <none>        8033/TCP,2112/TCP   ...

    Note: The configured metrics.port must be listed in the annotation prometheus.io/port. The default port is 2112.

  • Check if the ServiceMonitor resource with name modelmesh-metrics-monitor exists.

Additional troubleshooting steps can be found here.