Seldon/Iter8 - Progressive AB Test with Multiple Seldon Deployments

In this example we will AB Test two Iris models: an SKLearn model and an XGBOOST model. We will run a progressive rollout allowing Iter8 to control the traffic to the two Seldon Deployments and gradually move traffic to the best model.

Install Depenendcies

• Istio

• Seldon Core

• Seldon Core Analytics

• Iter8

You can create a Kind cluster with all dependencies installed with Ansible with:

pip install ansible openshift
ansible-galaxy collection install git+https://github.com/SeldonIO/ansible-k8s-collection.git,v0.1.0

Then from example/iter8 folder run:

ansible-playbook playbooks/iter8.yml

Create ABTest with Two Seldon Deployments

!cat baseline.yaml

apiVersion: v1
kind: Namespace
metadata:
    name: ns-baseline
---
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: iris
  namespace: ns-baseline
spec:
  predictors:
  - name: default
    graph:
      name: classifier
      modelUri: gs://seldon-models/v1.20.0-dev/sklearn/iris
      implementation: SKLEARN_SERVER

!kubectl apply -f baseline.yaml

namespace/ns-baseline created
seldondeployment.machinelearning.seldon.io/iris created

!cat candidate.yaml

apiVersion: v1
kind: Namespace
metadata:
    name: ns-candidate
---
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: iris
  namespace: ns-candidate
spec:
  predictors:
  - name: default
    graph:
      name: classifier
      modelUri: gs://seldon-models/xgboost/iris
      implementation: XGBOOST_SERVER

!kubectl apply -f candidate.yaml

namespace/ns-candidate created
seldondeployment.machinelearning.seldon.io/iris created

!kubectl wait --for condition=ready --timeout=600s pods --all -n ns-baseline

pod/iris-default-0-classifier-5dc67f64bf-brmss condition met

!kubectl wait --for condition=ready --timeout=600s pods --all -n ns-candidate

pod/iris-default-0-classifier-7fff869d67-g5qnh condition met

Create Virtual Service to Split Traffic

!cat routing-rule.yaml

apiVersion: networking.istio.io/v1alpha3
kind: VirtualService
metadata:
  name: routing-rule
  namespace: default
spec:
  gateways:
  - istio-system/seldon-gateway
  hosts:
  - iris.example.com
  http:
  - route:
    - destination:
        host: iris-default.ns-baseline.svc.cluster.local
        port:
          number: 8000
      headers:
        response:
          set:
            version: iris-v1
      weight: 100
    - destination:
        host: iris-default.ns-candidate.svc.cluster.local
        port:
          number: 8000
      headers:
        response:
          set:
            version: iris-v2
      weight: 0

!kubectl apply -f routing-rule.yaml

virtualservice.networking.istio.io/routing-rule created

Create some load on models.

We will send reqeusts which will be split by the Seldon AB Test as well as random feedback to both models with feedback favouring the candidate

!cat fortio.yaml

---
apiVersion: batch/v1
kind: Job
metadata:
  name: fortio-requests
  namespace: default
spec:
  template:
    spec:
      volumes:
      - name: shared
        emptyDir: {}    
      containers:
      - name: fortio
        image: fortio/fortio
        command: [ 'fortio', 'load', '-t', '6000s', '-qps', "5", '-json', '/shared/fortiooutput.json', '-H', 'Host: iris.example.com', '-H', 'Content-Type: application/json', '-payload', '{"data": {"ndarray":[[6.8,2.8,4.8,1.4]]}}',  "$(URL)" ]
        env:
        - name: URL
          value: URL_VALUE/api/v1.0/predictions
        volumeMounts:
        - name: shared
          mountPath: /shared         
      - name: busybox
        image: busybox:1.28
        command: ['sh', '-c', 'echo busybox is running! && sleep 6000']          
        volumeMounts:
        - name: shared
          mountPath: /shared       
      restartPolicy: Never
---
apiVersion: batch/v1
kind: Job
metadata:
  name: fortio-irisv1-rewards
  namespace: default
spec:
  template:
    spec:
      volumes:
      - name: shared
        emptyDir: {}    
      containers:
      - name: fortio
        image: fortio/fortio
        command: [ 'fortio', 'load', '-t', '6000s', '-qps', "0.7", '-json', '/shared/fortiooutput.json', '-H', 'Content-Type: application/json', '-payload', '{"reward": 1}',  "$(URL)" ]
        env:
        - name: URL
          value: URL_VALUE/seldon/ns-baseline/iris/api/v1.0/feedback
        volumeMounts:
        - name: shared
          mountPath: /shared         
      - name: busybox
        image: busybox:1.28
        command: ['sh', '-c', 'echo busybox is running! && sleep 6000']          
        volumeMounts:
        - name: shared
          mountPath: /shared       
      restartPolicy: Never
---
apiVersion: batch/v1
kind: Job
metadata:
  name: fortio-irisv2-rewards
  namespace: default
spec:
  template:
    spec:
      volumes:
      - name: shared
        emptyDir: {}    
      containers:
      - name: fortio
        image: fortio/fortio
        command: [ 'fortio', 'load', '-t', '6000s', '-qps', "1", '-json', '/shared/fortiooutput.json', '-H', 'Content-Type: application/json', '-payload', '{"reward": 1}',  "$(URL)" ]
        env:
        - name: URL
          value: URL_VALUE/seldon/ns-candidate/iris/api/v1.0/feedback
        volumeMounts:
        - name: shared
          mountPath: /shared         
      - name: busybox
        image: busybox:1.28
        command: ['sh', '-c', 'echo busybox is running! && sleep 6000']          
        volumeMounts:
        - name: shared
          mountPath: /shared       
      restartPolicy: Never

!URL_VALUE="http://$(kubectl -n istio-system get svc istio-ingressgateway -o jsonpath='{.spec.clusterIP}')" && \
  sed "s+URL_VALUE+${URL_VALUE}+g" fortio.yaml | kubectl apply -f -

job.batch/fortio-requests created
job.batch/fortio-irisv1-rewards created
job.batch/fortio-irisv2-rewards created

!kubectl wait --for condition=ready --timeout=600s pods --all -n default

pod/fortio-irisv1-rewards-t5drl condition met
pod/fortio-irisv2-rewards-rb9k8 condition met
pod/fortio-requests-fkp95 condition met

Create Metrics to evaluate

These are a standard set of metrics we use in all examples.

!cat ../../metrics.yaml

apiVersion: v1
kind: Namespace
metadata:
  name: iter8-seldon
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: 95th-percentile-tail-latency
  namespace: iter8-seldon
spec:
  description: 95th percentile tail latency
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      histogram_quantile(0.95, sum(rate(seldon_api_executor_client_requests_seconds_bucket{seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) by (le))
  provider: prometheus
  sampleSize: iter8-seldon/request-count
  type: Gauge
  units: milliseconds
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: error-count
  namespace: iter8-seldon
spec:
  description: Number of error responses
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      sum(increase(seldon_api_executor_server_requests_seconds_count{code!='200',seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0)
  provider: prometheus
  type: Counter
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query  
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: error-rate
  namespace: iter8-seldon
spec:
  description: Fraction of requests with error responses
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      (sum(increase(seldon_api_executor_server_requests_seconds_count{code!='200',seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0)) / (sum(increase(seldon_api_executor_server_requests_seconds_count{seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0))
  provider: prometheus
  sampleSize: iter8-seldon/request-count
  type: Gauge
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query    
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: mean-latency
  namespace: iter8-seldon
spec:
  description: Mean latency
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      (sum(increase(seldon_api_executor_client_requests_seconds_sum{seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0)) / (sum(increase(seldon_api_executor_client_requests_seconds_count{seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0))
  provider: prometheus
  sampleSize: iter8-seldon/request-count
  type: Gauge
  units: milliseconds
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query      
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: request-count
  namespace: iter8-seldon
spec:
  description: Number of requests
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      sum(increase(seldon_api_executor_client_requests_seconds_sum{seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0)
  provider: prometheus
  type: Counter
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query
---
apiVersion: iter8.tools/v2alpha2
kind: Metric
metadata:
  name: user-engagement
  namespace: iter8-seldon
spec:
  description: Number of feedback requests
  jqExpression: .data.result[0].value[1] | tonumber
  params:
  - name: query
    value: |
      sum(increase(seldon_api_executor_server_requests_seconds_count{service='feedback',seldon_deployment_id='$sid',predictor_name='$predictor',kubernetes_namespace='$ns'}[${elapsedTime}s])) or on() vector(0)
  provider: prometheus
  type: Gauge
  urlTemplate: http://seldon-core-analytics-prometheus-seldon.seldon-system/api/v1/query

!kubectl create -f ../../metrics.yaml

namespace/iter8-seldon created
metric.iter8.tools/95th-percentile-tail-latency created
metric.iter8.tools/error-count created
metric.iter8.tools/error-rate created
metric.iter8.tools/mean-latency created
metric.iter8.tools/request-count created
metric.iter8.tools/user-engagement created

!kubectl get metrics -n iter8-seldon

NAME                           TYPE      DESCRIPTION
95th-percentile-tail-latency   Gauge     95th percentile tail latency
error-count                    Counter   Number of error responses
error-rate                     Gauge     Fraction of requests with error responses
mean-latency                   Gauge     Mean latency
request-count                  Counter   Number of requests
user-engagement                Gauge     Number of feedback requests

Create Progressive Rollout Experiment

• Run 15 iterations with 5 second gaps between default and candidate models

• Both models must pass objectives

• winnder will be chosen based on user engagement metric

!cat experiment.yaml

apiVersion: iter8.tools/v2alpha2
kind: Experiment
metadata:
  name: quickstart-exp
spec:
  target: iris
  strategy:
    testingPattern: A/B
    deploymentPattern: Progressive
    actions:
      # when the experiment completes, promote the winning version using kubectl apply
      finish:
      - task: common/exec
        with:
          cmd: /bin/bash
          args: [ "-c", "kubectl apply -f {{ .promote }}" ]
  criteria:
    requestCount: iter8-seldon/request-count
    rewards: # Business rewards
    - metric: iter8-seldon/user-engagement
      preferredDirection: High # maximize user engagement
    objectives:
    - metric: iter8-seldon/mean-latency
      upperLimit: 2000
    - metric: iter8-seldon/95th-percentile-tail-latency
      upperLimit: 5000
    - metric: iter8-seldon/error-rate
      upperLimit: "0.01"
  duration:
    intervalSeconds: 10
    iterationsPerLoop: 10
  versionInfo:
    # information about model versions used in this experiment
    baseline:
      name: iris-v1
      weightObjRef:
        apiVersion: networking.istio.io/v1alpha3
        kind: VirtualService
        name: routing-rule
        namespace: default
        fieldPath: .spec.http[0].route[0].weight      
      variables:
      - name: ns
        value: ns-baseline
      - name: sid
        value: iris
      - name: predictor
        value: default
      - name: promote
        value: https://raw.githubusercontent.com/iter8-tools/iter8/master/samples/seldon/quickstart/promote-v1.yaml
    candidates:
    - name: iris-v2
      weightObjRef:
        apiVersion: networking.istio.io/v1alpha3
        kind: VirtualService
        name: routing-rule
        namespace: default
        fieldPath: .spec.http[0].route[1].weight      
      variables:
      - name: ns
        value: ns-candidate
      - name: sid
        value: iris
      - name: predictor
        value: default
      - name: promote
        value: https://raw.githubusercontent.com/iter8-tools/iter8/master/samples/seldon/quickstart/promote-v2.yaml

!kubectl create -f experiment.yaml

experiment.iter8.tools/quickstart-exp created

Monitor Experiment

Download iter8ctl.

GO111MODULE=on GOBIN=/usr/local/bin go get github.com/iter8-tools/[email protected]

Then:

while clear; do kubectl get experiment quickstart-exp -o yaml | iter8ctl describe -f -; sleep 8; done

By the end you should see the xgboost candidate model is promoted.

!kubectl wait experiment quickstart-exp --for=condition=Completed --timeout=300s

experiment.iter8.tools/quickstart-exp condition met

!kubectl get experiment quickstart-exp

NAME             TYPE   TARGET   STAGE       COMPLETED ITERATIONS   MESSAGE
quickstart-exp   A/B    iris     Completed   10                     ExperimentCompleted: Experiment Completed

Cleanup

!kubectl delete -f fortio.yaml
!kubectl delete -f experiment.yaml
!kubectl delete -f ../../metrics.yaml
!kubectl delete -f routing-rule.yaml
!kubectl delete -f baseline.yaml
!kubectl delete -f candidate.yaml

job.batch "fortio-requests" deleted
job.batch "fortio-irisv1-rewards" deleted
job.batch "fortio-irisv2-rewards" deleted
experiment.iter8.tools "quickstart-exp" deleted
namespace "iter8-seldon" deleted
metric.iter8.tools "95th-percentile-tail-latency" deleted
metric.iter8.tools "error-count" deleted
metric.iter8.tools "error-rate" deleted
metric.iter8.tools "mean-latency" deleted
metric.iter8.tools "request-count" deleted
metric.iter8.tools "user-engagement" deleted
virtualservice.networking.istio.io "routing-rule" deleted
namespace "ns-baseline" deleted
seldondeployment.machinelearning.seldon.io "iris" deleted
namespace "ns-candidate" deleted
seldondeployment.machinelearning.seldon.io "iris" deleted