Skip to content

Latest commit

 

History

History
56 lines (42 loc) · 2.44 KB

OPERATIONS.md

File metadata and controls

56 lines (42 loc) · 2.44 KB
Raw

Procedures for Maintaining a Harmony Kubernetes Cluster

This document provides instructions for the ongoing maintenance of a running Harmony cluster.

Monitoring

Monitoring a running Harmony cluster is done using a combination of custom logging (sent to CloudWatch when run on AWS) and the tools described here.

Prometheus

The deploy-prometheus script in the bin directory of this repository can be used to deploy Prometheus into the Harmony Kubernetes cluster along with the kube-state-metrics service which provides useful metrics related to pods and the state of the cluster. The Prometheus UI can be accessed by forwarding the service's port (9090) to a local machine as follows (all kubectl commands listed in this document use either the default context or the one specified by the optional --kubeconfig paramter):

  1. get the pod identifier for Prometheus
kubectl get pods -n monitoring [--kubeconfig <config file>]
  1. forward port 9090 from the pod to your local machine
kubectl port-forward <Prometheus Pod ID> 9090:9090 -n monitoring [--kubeconfig <config file>]

After which you can access the Prometheus UI in web browser at http://localhost:9090.

Metrics

Harmony services workers provide a metric that is scraped by Prometheus, num_ready_work_items, which is the number of work items for a given service that are in the 'READY' state (i.e., available to be worked but not yet being worked). It can be accessed via the Prometheus UI.

Harmony deploys the Prometheus Adapter to make this metric available to the Horizontal Pod Autoscaler (HPA) to scale the number of pods for a given service up or down in response to load.

Monitoring the Number of Pods for Each Service

The following Prometheus query (promQL) can be used to get the current number of running pods for each deployed service.

sum(
count(kube_pod_status_phase{phase="Running", namespace="harmony"}) by (pod)
*
on (pod) group_left(label_name) sum (kube_pod_labels{namespace="harmony",label_name=~".*"}) without (namespace)
) by (label_name)

The ".*" for the label_name can be replaced with a service name to get the count for just that service, e.g, "harmony-service-example".