Optimizing a Kubernetes application

In this example, we’ll optimize Online Boutique, a demo e-commerce application running on microservices, by tuning the resources allocated to a selection of pods. This is a common use case where we want to minimize the cost associated with running an application without impacting the SLO.

Notice: all the required artifacts are published in this public repository.

Environment setup

The test environment includes the following instances:

• Akamas: the instance running Akamas.

• Cluster: an instance hosting a Minikube cluster.

You can configure the Minikube cluster using the scripts provided in the public repository by running the command

kubectl apply -f kubernetes-online-boutique/kube/prometheus.yaml

Telemetry Infrastructure setup

To gather metrics about the application we will use Prometheus. It will be automatically configured by applying the artifacts in the repository with the following command:

kubectl apply -f kubernetes-online-boutique/kube/

Application and Test tool

The targeted system is Online Boutique, a microservice-based demo application. In the same namespace, a deployment running the load generator will stress the boutique and forward the performance metrics to Prometheus.

To configure the application and the load generator on your (Minikube) cluster, apply the definitions provided in the public repository by running the following command:

kubectl apply -f kubernetes-online-boutique/kube/

Optimization setup

In this section, we will guide you through the steps required to set up the optimization on Akamas.

If you have not installed the Kubernetes optimization pack yet, take a look at the Kubernetes optimization pack page to proceed with the installation.

Notice: the artifacts to create the Akamas entities can be found in the public repository, under the akamas directory.

System

System Online Boutique

Here’s the definition of the system containing our components and telemetry-instances for this example:

name: Online Boutique
description: The Online Boutique by Google

To create the system run the following command:

akamas create component application.yaml 'Online Boutique'

Component online_boutique

We’ll use a component of type WebApplication to represent at a high level the Online Boutique application. To identify the related Prometheus metrics the configuration requires the prometheus property for the telemetry service, detailed later in this guide.

Here’s the definition of the component:

name: online_boutique
description: The Online Boutique application
componentType: Web Application
properties:
  prometheus:
    instance: .*
    job: .*
    namespace: akamas-demo
    container: server|redis

To create the component in the system run the following command:

akamas create component application.yaml 'Online Boutique'

Component frontend and productcatalogservice

The public repository contains the definition of all the services that compose Online Boutique. In this guide, for the sake of simplicity, we’ll only tune the resources of the containers in the frontend and the product-catalog pods, defined as components of type Kubernetes Container.

Here’s their definition:

name: frontend
description: The frontend of the online boutique by Google
componentType: Kubernetes Container
properties:
  prometheus:
    job: .*
    instance: .*
    name: .*
    pod: ak-frontend.*
    container: server

name: productcatalogservice
description: The productcatalogservice of the online boutique by Google
componentType: Kubernetes Container
properties:
  prometheus:
    job: .*
    instance: .*
    name: .*
    pod: ak-productcatalogservice.*
    container: server

To create the component in the system run the following command:

akamas create component frontend.yaml 'Online Boutique'
akamas create component productcatalogservice.yaml 'Online Boutique'

Workflow

The workflow is divided into the following steps:

• Create the YAML artifacts with the updated resource limits for the tuned containers.

• Apply the updated definitions to the cluster.

• Wait for the rollout to complete.

• Start the load generator

• Let the test run for a fixed amount of time

• Stop the test and reset the load generator

The following is the definition of the workflow:

name: boutique
tasks:
  - name: Configure Online Boutique
    operator: FileConfigurator
    arguments:
      source:
        hostname: CLUSTER_INSTANCE_IP
        username: akamas
        password: akamas
        path: boutique.yaml.templ
      target:
        hostname: cluster
        username: akamas
        password: akamas
        path: boutique.yaml

  - name: Apply new configuration to the Online Boutique
    operator: Executor
    arguments:
      host:
        hostname: CLUSTER_INSTANCE_IP
        username: akamas
        password: akamas
      command: kubectl apply -f boutique.yaml

  - name: Check Online Boutique is up
    operator: Executor
    arguments:
      retries: 0
      host:
        hostname: CLUSTER_INSTANCE_IP
        username: akamas
        password: akamas
      command: kubectl wait --for=condition=available deploy/ak-frontend deploy/ak-productcatalogservice --timeout=30s

  - name: Start Locust Test
    operator: Executor
    arguments:
      host:
        hostname: CLUSTER_INSTANCE_IP
        username: akamas
        password: akamas
      command: bash load-test.sh start

  - name: Test
    operator: Sleep
    arguments:
      seconds: 150

  - name: Stop Locust test
    operator: Executor
    arguments:
      host:
        hostname: CLUSTER_INSTANCE_IP
        username: akamas
        password: akamas
      command: bash load-test.sh stop

To better illustrate the process, here is a snippet of the template file used to update the resource limits for the frontend deployment.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: ak-frontend
  namespace: akamas-demo
spec:
  selector:
    matchLabels:
      app: ak-frontend
  template:
    metadata:
      labels:
        app: ak-frontend
    # other definitions...
    spec:
      containers:
        - name: server
          image: gcr.io/google-samples/microservices-demo/frontend:v0.2.2
          # other definitions...
          resources:
            requests:
              cpu: ${frontend.cpu_limit}
              memory: ${frontend.memory_limit}
            limits:
              cpu: ${frontend.cpu_limit}
              memory: ${frontend.memory_limit}
# other definitions...

The following are respectively the script to start and stop the load generator:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: ak-frontend
  namespace: akamas-demo
spec:
  selector:
    matchLabels:
      app: ak-frontend
  replicas: 1
  strategy:
    rollingUpdate:
      maxSurge: 1
      maxUnavailable: 1
  template:
    metadata:
      labels:
        app: ak-frontend
    spec:
      serviceAccountName: default
      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 1
              preference:
                matchExpressions:
                  - key: akamas/node
                    operator: In
                    values:
                      - akamas
      containers:
        - name: server
          image: gcr.io/google-samples/microservices-demo/frontend:v0.2.2
          ports:
            - containerPort: 8080
          readinessProbe:
            initialDelaySeconds: 10
            httpGet:
              path: "/_healthz"
              port: 8080
              httpHeaders:
                - name: "Cookie"
                  value: "shop_session-id=x-readiness-probe"
          livenessProbe:
            initialDelaySeconds: 10
            httpGet:
              path: "/_healthz"
              port: 8080
              httpHeaders:
                - name: "Cookie"
                  value: "shop_session-id=x-liveness-probe"
          env:
            - name: PORT
              value: "8080"
            - name: PRODUCT_CATALOG_SERVICE_ADDR
              value: "ak-productcatalogservice:3550"
            - name: CURRENCY_SERVICE_ADDR
              value: "ak-currencyservice:7000"
            - name: CART_SERVICE_ADDR
              value: "ak-cartservice:7070"
            - name: RECOMMENDATION_SERVICE_ADDR
              value: "ak-recommendationservice:8080"
            - name: SHIPPING_SERVICE_ADDR
              value: "ak-shippingservice:50051"
            - name: CHECKOUT_SERVICE_ADDR
              value: "ak-checkoutservice:5050"
            - name: AD_SERVICE_ADDR
              value: "ak-adservice:9555"
            - name: ENV_PLATFORM
              value: "aws"
            - name: DISABLE_TRACING
              value: "1"
            - name: DISABLE_PROFILER
              value: "1"
          resources:
            requests:
              cpu: ${frontend.cpu_limit}
              memory: ${frontend.memory_limit}
            limits:
              cpu: ${frontend.cpu_limit}
              memory: ${frontend.memory_limit}

---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: ak-productcatalogservice
  namespace: akamas-demo
spec:
  selector:
    matchLabels:
      app: ak-productcatalogservice
  replicas: 1
  template:
    metadata:
      labels:
        app: ak-productcatalogservice
    spec:
      serviceAccountName: default
      terminationGracePeriodSeconds: 5
      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 1
              preference:
                matchExpressions:
                  - key: akamas/node
                    operator: In
                    values:
                      - akamas
      containers:
        - name: server
          image: gcr.io/google-samples/microservices-demo/productcatalogservice:v0.2.2
          ports:
            - containerPort: 3550
          env:
            - name: PORT
              value: "3550"
            - name: DISABLE_STATS
              value: "1"
            - name: DISABLE_TRACING
              value: "1"
            - name: DISABLE_PROFILER
              value: "1"
          readinessProbe:
            exec:
              command: ["/bin/grpc_health_probe", "-addr=:3550"]
          livenessProbe:
            exec:
              command: ["/bin/grpc_health_probe", "-addr=:3550"]
          resources:
            requests:
              cpu: ${productcatalogservice.cpu_limit}
              memory: ${productcatalogservice.memory_limit}
            limits:
              cpu: ${productcatalogservice.cpu_limit}
              memory: ${productcatalogservice.memory_limit}

#/bin/bash

ACTION=$1

LOCUST_ENDPOINT="$(minikube service -n akamas-demo ak-loadgenerator | awk '/web-ui.*http/ {print $8}')"

case $ACTION in
  start)
    echo curl -X POST -d 'user_count=100' -d 'spawn_rate=3' -d 'host=http://ak-frontend:80'  "${LOCUST_ENDPOINT}/swarm"
    ;;
  stop)
    echo curl "${LOCUST_ENDPOINT}/stop"
    echo curl "${LOCUST_ENDPOINT}/stats/reset"
    ;;
  *)
    echo "Unrecognized option '${ACTION}'"
    exit 1
    ;;
esac

Telemetry

If you have not installed the Prometheus telemetry provider yet, take a look at the telemetry provider page Prometheus provider to proceed with the installation.

With the definition of the telemetry instance shown below, we import the end-user performance metrics provided by the load-generator, along with a custom definition of "cost" given by a weighted sum of the CPU and memory allocated for the pods in the cluster:

provider: Prometheus
config:
  address: CLUSTER_IP
  port: PROM_PORT
metrics:
  - metric: users
    datasourceMetric: "locust_users"
  - metric: transactions_throughput
    datasourceMetric: 'rate(locust_requests_num_requests{name="Aggregated"}[30s]) - rate(locust_requests_num_failures{name="Aggregated"}[30s])'
  - metric: transactions_error_throughput
    datasourceMetric: 'rate(locust_requests_num_failures{name="Aggregated"}[30s])'
  - metric: transactions_error_rate
    datasourceMetric: "locust_requests_fail_ratio"
  - metric: transactions_response_time
    datasourceMetric: 'locust_requests_avg_response_time{name="Aggregated"}'
  - metric: transactions_response_time_p50
    datasourceMetric: 'locust_requests_current_response_time_percentile_50'
  - metric: transactions_response_time_p95
    datasourceMetric: 'locust_requests_current_response_time_percentile_95'

  - metric: cost
    datasourceMetric: 'sum(kube_pod_container_resource_requests{resource="cpu" %FILTERS%})*29 + sum(kube_pod_container_resource_requests{resource="memory" %FILTERS%})/1024/1024/1024*3.2'

To create the telemetry instance execute the following command:

akamas create telemetry-instance prometheus.yml 'Online Boutique'

Study

With this study, we want to minimize the "cost" of running the application, which, according to the definition described in the previous section, it means reducing the resources allocated to the tuned pods in the cluster. At the same time, we want the application to stay within the expected SLO, and that is obtained by defining a constraint on the response time and error rate recorded by the load generator.

name: Minimize Kubernetes Online Boutique cost while matching SLOs
system: Online Boutique
workflow: boutique

goal:
  objective: minimize
  constraints:
    absolute:
      - name: response_time
        formula: online_boutique.transactions_response_time <= 500
      - name: error_rate
        formula: online_boutique.transactions_error_rate <= 0.02
  function:
    formula: online_boutique.cost

windowing:
  type: trim
  trim: [1m, 30s]
  task: Test

metricsSelection:
  - online_boutique.cost
  - online_boutique.transactions_throughput
  - online_boutique.transactions_error_rate
  - online_boutique.transactions_response_time
  - online_boutique.transactions_response_time_p95
  - online_boutique.users
  - frontend.container_cpu_used
  - frontend.container_cpu_util
  - frontend.container_cpu_limit
  - frontend.container_cpu_throttle_time
  - frontend.container_memory_used
  - frontend.container_memory_util
  - frontend.container_memory_limit
  - productcatalogservice.container_cpu_used
  - productcatalogservice.container_cpu_util
  - productcatalogservice.container_cpu_limit
  - productcatalogservice.container_cpu_throttle_time
  - productcatalogservice.container_memory_used
  - productcatalogservice.container_memory_util
  - productcatalogservice.container_memory_limit

parametersSelection:
  - name: frontend.cpu_limit
    domain: [100, 300]
  - name: frontend.memory_limit
    domain: [64, 512]
  - name: productcatalogservice.cpu_limit
    domain: [100, 500]
  - name: productcatalogservice.memory_limit
    domain: [64, 512]

steps:
  - name: baseline
    type: baseline
    values:
      frontend.cpu_limit: 300
      frontend.memory_limit: 256
      productcatalogservice.cpu_limit: 300
      productcatalogservice.memory_limit: 256

  - name: optimize
    type: optimize
    numberOfExperiments: 50

To create and run the study execute the following commands:

akamas create study study.yaml
akamas start study 'Minimize Kubernetes Online Boutique cost while matching SLOs'