Seldon Core exposes metrics that can be scraped by Prometheus. The core metrics are exposed by the service orchestrator (executor).

The metrics are:

Prediction Requests

  • Requests to the service orchestrator from an ingress, e.g. API gateway or Ambassador

    • seldon_api_executor_server_requests_seconds_(bucket,count,sum) - histogram type metric

    • seldon_api_executor_server_requests_seconds_summary_(count,sum) - summary type metric

  • Requests from the service orchestrator to a component, e.g., a model

    • seldon_api_executor_client_requests_seconds_(bucket,count,sum) - histogram type metric

    • seldon_api_executor_client_requests_seconds_summary_(count,sum) - summary type metric

Each metric has the following key value pairs for further filtering which will be taken from the SeldonDeployment custom resource that is running:

  • service

  • deployment_name

  • predictor_name

  • predictor_version (This will be derived from the predictor metadata labels)

  • model_name

  • model_image

  • model_version

Metrics with Prometheus Operator


We recommend to configure Prometheus using Prometheus Operator. The kube-prometheus stack configuration can be easily installed using the Bitnami Helm Charts

kubectl create namespace seldon-monitoring

helm upgrade --install seldon-monitoring kube-prometheus \
    --version 8.3.2 \
    --set fullnameOverride=seldon-monitoring \
    --namespace seldon-monitoring \

kubectl rollout status -n seldon-monitoring statefulsets/prometheus-seldon-monitoring-prometheus

The following pods should now be present in the seldon-monitoring namespace:

$ kubectl get pods -n seldon-monitoring
NAME                                            READY   STATUS    RESTARTS   AGE
alertmanager-seldon-monitoring-alertmanager-0   2/2     Running   0          51s
prometheus-kube-state-metrics-d97b6b5ff-n5z7w   1/1     Running   0          52s
prometheus-node-exporter-jmffw                  1/1     Running   0          52s
prometheus-seldon-monitoring-prometheus-0       2/2     Running   0          51s
seldon-monitoring-operator-6d558f5696-xhq66     1/1     Running   0          52s


Following PodMonitor resource will instruct Prometheus to scrape ports named metrics from pods managed by Seldon Core. Create seldon-podmonitor.yaml file

kind: PodMonitor
  name: seldon-podmonitor
  namespace: seldon-monitoring
    matchLabels: seldon-core
    - port: metrics
      path: /prometheus
    any: true

and apply it with

kubectl apply -f seldon-podmonitor.yaml


Assuming that there exist SeldonDeployment models running in the cluster one can verify Prometheus metrics by accessing the Prometheus UI.

Expose Prometheus to your localhost with

$ kubectl port-forward -n seldon-monitoring svc/seldon-monitoring-prometheus 9090:9090

You can now head to your browser http://localhost:9090 to access the Prometheus UI. Start by verifying at Status -> Targets information.


Then, head to Graph section and query for seldon_api_executor_client_requests_seconds_count. You should see output similar to following (assuming that your SeldonDeployments are receiving some inference requests)


Custom Metrics

Seldon Core exposes basic metrics via Prometheus endpoints on its service orchestrator that include request count, request time percentiles and rolling accuracy for each running model as described in metrics documentation. However, you may wish to expose custom metrics from your components which are automatically added to Prometheus. For this purpose you can supply extra fields in the returned meta data of the response object in the API calls to your components as illustrated below:

    "meta": {
        "metrics": [
                "type": "COUNTER",
                "key": "mycounter",
                "value": 1.0,
                "tags": {"mytag": "mytagvalue"}
                "type": "GAUGE",
                "key": "mygauge",
                "value": 22.0
                "type": "TIMER",
                "key": "mytimer",
                "value": 1.0
    "data": {
        "ndarray": [

We provide three types of metric that can be returned in the meta.metrics list:

  • COUNTER : a monotonically increasing value. It will be added to any existing value from the metric key.

  • GAUGE : an absolute value showing a level, it will overwrite any existing value.

  • TIMER : a time value (in msecs), it will be aggregated into Prometheus’ HISTOGRAM.

Each metric, apart from the type, takes a key and a value. The proto buffer definition is shown below:

message Metric {
 enum MetricType {
     COUNTER = 0;
     GAUGE = 1;
     TIMER = 2;
 string key = 1;
 MetricType type = 2;
 float value = 3;
 map<string,string> tags = 4;

Metrics endpoints

Custom metrics are exposed directly by the Python wrapper. In order for Prometheus to scrape multiple endpoints from a single Pod we use metrics name for ports that expose Prometheus metrics:

- containerPort: 6000
  name: metrics
  protocol: TCP

If you configured Prometheus using Prometheus Operator as discussed above you should be set to go. If, however, you are configuring Prometheus not using Prometheus Operator this require us to use a following entry

- source_labels: [__meta_kubernetes_pod_container_port_name]
  action: keep
  regex: metrics(-.*)?

in the Prometheus config together with following two annotations: "true" "/prometheus"

Note: we do not use annotation in this configuration.

Before Seldon Core 1.1 custom metrics have been returned to the orchestrator which exposed them all together to Prometheus via a single endpoint. We used to have at this time all three following annotations: "true" "/prometheus" "8000"


As we expose the metrics via Prometheus, if tags are added they must appear in every metric response otherwise Prometheus will consider such metrics as a new time series, see official documentation.

Before Seldon Core 1.1 orchestrator enforced presence of same set of labels using the micrometer library to expose metrics. Exceptions would happen if this condition have been violated.

Supported wrappers

At present the following Seldon Core wrappers provide integrations with custom metrics:


There is an example notebook you can use to test the metrics.