Kubernetes Redis Queue HPA

This repo demonstrates using a Kubernetes custom metric to enable the Kubernetes Horizontal Pod Autoscaler (HPA) to scale based on redis queue (rq) length plus busy worker count.

NOTE: If you want to scale on CPU and/or memory, there's no need for this repo or all this complexity. Use the Metrics Server instead.

Installing with helm

helm install suihei . -n YOUR_NAMESPACE_HERE --create-namespace -f values.yaml -f values-custom.yaml -f values-custom-secret.yaml

Uninstalling with helm

helm uninstall suihei

Kubernetes custom metrics

• List all metric definitions:

  kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1/" | jq .

  Sample output:

  {
    "kind": "APIResourceList",
    "apiVersion": "v1",
    "groupVersion": "custom.metrics.k8s.io/v1beta1",
    "resources": [
      {
        "name": "deployments.apps/redisqueue_length",
        "singularName": "",
        "namespaced": false,
        "kind": "MetricValueList",
        "verbs": ["get"]
      }
    ]
  }

• Get metric values for one metric

  namespace=YOUR_NAMESPACE_HERE
  kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1/namespaces/$namespace/deployments.apps/rq-worker/redisqueue_length?metricLabelSelector=queues%3Dhigh-low" | jq .

  Sample output (see also MetricValueList Golang struct and protobuf message):

  {
    "apiVersion": "custom.metrics.k8s.io/v1beta1",
    "items": [
      {
        "_debug": {
          "all_busy_workers": [
            "26db71be1f254f99b718a2cfa2219145",
            "8ba55f15a5a344968ca35bbdb0a44461"
          ],
          "queue_info": {
            "high": {
              "busy_workers": [
                "8ba55f15a5a344968ca35bbdb0a44461"
              ],
              "queued_job_count": 6
            },
            "low": {
              "busy_workers": [
                "26db71be1f254f99b718a2cfa2219145"
              ],
              "queued_job_count": 1
            }
          }
        },
        "describedObject": {
          "apiVersion": "apps/v1",
          "kind": "Deployment",
          "name": "rq-worker",
          "namespace": "YOUR_NAMESPACE_HERE"
        },
        "metricName": "count",
        "timestamp": "2022-02-04T11:55:41Z",
        "value": "9"
      }
    ],
    "kind": "MetricValueList",
    "metadata": {
      "selflink": "/apis/custom.metrics.k8s.io/v1beta1/"
    }
  }

Ports

• 32000: apiserver (see .customMetrics.apiServer.port in values.yaml)

URLs

Use ip="$(minikube ip)" to get the IP address of the cluster.

• Enqueue job:

  queue=high
  curl -XPOST \
    -H 'Content-Type: application/json' \
    -d '{"sleep": 60}' \
    "http://$ip:32000/queues/$queue/enqueue"

  Sample output:

  {
    "job": {
      "id": "uuiduuid-uuid-4uid-uuid-uuiduuiduuid",
      "result": null,
      "status": "queued"
    }
  }

Troubleshooting

Check the metrics API service is working:

kubectl get apiservice | grep custom.metrics

v1beta1.custom.metrics.k8s.io  YOUR_NAMESPACE_HERE/custom-metrics-apiserver  True  1h

If kubernetes won't allow helm to reinstall the app due to lingering objects after an uninstall, use the following:

kubectl proxy &  # get a proxy listening on port 8001.
namespace=YOUR_NAMESPACE_HERE
curl -k \
  -XPUT \
  -H "Content-Type: application/json" \
  --data-binary '{"apiVersion":"v1","kind":"Namespace","metadata":{"name":"'"$namespace"'"},"spec":{"finalizers":[]}}' \
  "http://127.0.0.1:8001/api/v1/namespaces/$namespace/finalize"

License

See LICENSE and licensing.md for details.