Canary Release com Docker, Kubernetes, Gitlab e Traefik

#startup #docker #k8s #traefik #gitlab #ci/cd #helm #kops #aws #help

About me

From Fortaleza - CE

10 years+ coding

CTO/Co-founder @Agendor

Java, Javascript, Android, PHP, Node.js, Ruby, React, Infra, Linux, AWS, MySQL, Business, Product.

Tulio Monte Azul

@tuliomonteazul

tulio@agendor.com.br

Agendor é a ferramenta principal do vendedor.

(Github do Vendedor)

Versão Web e Mobile com plano gratuito e pago.

O que esperar dessa talk?

"o quê"  vs  "como"

"release"  vs  "implementação"

Agenda

  • Contexto
  • Motivação
  • Estratégia
  • Desafios
  • Lições aprendidas

Contexto

  • Projeto em PHP (legado)
  • Migração pra SPA com Rails + React
  • Equipe enxuta (6 devs)
  • Mercado competitivo
  • Churn (desistências)
  • 25k usuários ativos
  • 5k rpm

Motivação

Problemas

Solução

  • Deploy com bug = caos

  • Medo de arriscar e atualizar versões

  • Processos manuais (ssh, yum install)

  • Ambiente dev != Outro dev != Prod

  • Falta de testes automatizados

  • Downtime

  • Improdutividade

  • Deploy com rollback

  • Deploy com ambientes isolados (containers)

  • Infrastructure as Code
     

  • Infrastructure as Code
    ​

  • Canary Release

  • Monitoramento e auto recover

  • CI/CD

Soluções

CI/CD 

Rollback fácil

muito fácil

Automação

Custo

stages:
  - test
  - release
  - deploy

image: azukiapp/docker-kubernetes-ci:17.04

services:
- docker:dind

variables:
  DOCKER_DRIVER: overlay

before_script:
  - docker info
  - docker-compose --version

  # Install kubectl and generate configurations
  # More info in https://gitlab.com/agendorteam/agendor-infra#configure-kubectl-context-in-ci
  - kube-config-generator
  - kubectl cluster-info
  - helm version

# Test
###########
test:
  stage: test
  tags:
    - docker

  script:
    - docker-compose build test
    - docker-compose run test

# Release
###########
.release: &release
  stage: release
  script:
    # REFERENCES: https://docs.gitlab.com/ce/ci/variables/
    # - ./scripts/release.sh agendor-web master
    - ./scripts/release.sh $CI_PROJECT_NAME $IMAGE_TAG

release_current_staging:
  <<: *release
  tags:
    - docker
  variables:
    DEPLOY_TYPE: staging
    IMAGE_TAG: stg-$CI_BUILD_REF_SLUG-$CI_BUILD_REF
  except:
    - master

release_current_test:
  <<: *release
  tags:
    - docker
  variables:
    DEPLOY_TYPE: test
    IMAGE_TAG: tst-$CI_BUILD_REF_SLUG-$CI_BUILD_REF
  except:
    - master

release_current_production:
  <<: *release
  tags:
    - docker
  variables:
    DEPLOY_TYPE: production
    IMAGE_TAG: $CI_BUILD_REF_SLUG-$CI_BUILD_REF
  only:
    - master

release_latest_production:
  <<: *release
  tags:
    - docker
  variables:
    DEPLOY_TYPE: production
    IMAGE_TAG: latest
  only:
    - master

# Deploy
###########
.deploy_template: &deploy_definition
  stage: deploy
  tags:
    - docker
  script:
    # ./scripts/deploy.sh [PROJECT_NAME [REF_NAME [HELM_NAME]]]
    - ./scripts/deploy.sh $CI_PROJECT_NAME $IMAGE_TAG
    - ./scripts/deploy.sh $CI_PROJECT_NAME $IMAGE_TAG agendor-api-mobile

deploy_staging:
  <<: *deploy_definition
  environment:
    name: staging
  variables:
    DEPLOY_TYPE: staging
    IMAGE_TAG: stg-$CI_BUILD_REF_SLUG-$CI_BUILD_REF
  only:
    - develop

deploy_canary:
  <<: *deploy_definition
  environment:
    name: canary
  variables:
    DEPLOY_TYPE: canary
    IMAGE_TAG: $CI_BUILD_REF_SLUG-$CI_BUILD_REF
  only:
    - master

# manuals
deploy_staging_manual:
  <<: *deploy_definition
  environment:
    name: staging
  variables:
    DEPLOY_TYPE: staging
    IMAGE_TAG: stg-$CI_BUILD_REF_SLUG-$CI_BUILD_REF
  when: manual
  except:
    - develop
    - master

deploy_test_manual:
  <<: *deploy_definition
  environment:
    name: test
  variables:
    DEPLOY_TYPE: test
    IMAGE_TAG: tst-$CI_BUILD_REF_SLUG-$CI_BUILD_REF
  when: manual
  except:
    - master

deploy_production_manual:
  <<: *deploy_definition
  environment:
    name: production
  variables:
    DEPLOY_TYPE: production
  when: manual
  only:
    - master
gitlab-ci.yaml

Soluções

Usado em produção

Google, Red Hat, CoreOs

Monitoramento, Rolling Updates, Auto recovery

Comunidade

Documentação

Soluções

Usado em produção

Múltiplos ambientes

Deploy por ambiente

Hipster

Infrastructure as Code

Ambiente dev = prod

Host mapping

Fast

Less code = less bug

Canary Release

Soluções

frontends:
  # agendor-web
  web:
    rules:
      - "Host:web.agendor.com.br"
  web_canary:
    rules:
      - "Host:web.agendor.com.br"
      - "HeadersRegexp:Cookie,_AUSER_TYPE=canary"
backends:
  ## agendor-web
  web:
    - name: k8s
      url: "http://agendor-web.production.svc.cluster.local"
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"

Estratégia

Estratégia - Fase 1

web.agendor.com.br

production-legacy

web.agendor.com.br
cookie: canary

production-canary

web.agendor.com.br

Estratégia - Fase 1

frontends:
  web:
    rules:
      - "Host:web.agendor.com.br"
  web_canary:
    rules:
      - "Host:web.agendor.com.br"
      - "HeadersRegexp:Cookie,_AUSER_TYPE=canary"
backends:
  web:
    - name: legacy_1
      url: http://ec2-54-10-XXX-XXX.sa-east-1.compute.amazonaws.com
    - name: legacy_2
      url: http://ec2-54-20-XXX-XXX.sa-east-1.compute.amazonaws.com
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"
aws.yaml
helm upgrade --kube-context k8s.agendor.com.br agendor-infra -f aws.yaml ./

Estratégia - Fase 2

web.agendor.com.br

production-legacy

web.agendor.com.br
cookie: canary

production-canary

90%

web.agendor.com.br

10%

Estratégia - Fase 2

backends:
  web:
    - name: legacy_1
      values:
        - weight = 5
        - url = "http://ec2-54-10-XXX-XXX.sa-east-1.compute.amazonaws.com"
    - name: legacy_2
      values:
        - weight = 5
        - url = "http://ec2-54-20-XXX-XXX.sa-east-1.compute.amazonaws.com"
    - name: k8s
      values:
        - weight = 1
        - url = "http://agendor-web-canary.production.svc.cluster.local"
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"
aws.yaml
helm upgrade --kube-context k8s.agendor.com.br agendor-infra -f aws.yaml ./

Estratégia - Fase 3

web.agendor.com.br

production

web.agendor.com.br
cookie: canary

production-canary

production-legacy

Estratégia - Fase 3

backends:
  web:
    - name: k8s
      url: "http://agendor-web.production.svc.cluster.local"
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"
aws.yaml
helm upgrade --kube-context k8s.agendor.com.br agendor-infra -f aws.yaml ./

Desafios

Pods Scaling

NewRelic FTW

# back-end
web_canary:
  - name: k8s
    url: "http://agendor-web-canary.production.svc.cluster.local"




# front-end
web_canary:
  rules:
    - "Host:web.agendor.com.br"
    - "HeadersRegexp:Cookie,_AUSER_TYPE=canary"
# back-end
web_canary:
  - name: k8s
    url: "http://agendor-web-canary.production.svc.cluster.local"
web_canary_host:
  - name: k8s
    url: "http://agendor-web-canary.production.svc.cluster.local"

# front-end
web_canary:
  rules:
    - "Host:web.agendor.com.br"
    - "HeadersRegexp:Cookie,_AUSER_TYPE=canary"
web_canary_host:
  rules:
    - "Host:web-canary.agendor.com.br"

Kubectl FTW

$ kubectl --namespace=production scale --replicas=8 deployment/agendor-web-canary
$ kubectl --namespace=production get pods -l=app=agendor-web-canary
NAME                                  READY     STATUS    RESTARTS   AGE
agendor-web-canary-1388324162-2qqgr   1/1       Running   0          5h
agendor-web-canary-1388324162-38xlb   1/1       Running   0          1m
agendor-web-canary-1388324162-3g815   1/1       Running   0          5h
agendor-web-canary-1388324162-89j6k   1/1       Running   0          5h
$ kubectl --namespace=production get pods -l=app=agendor-web-canary
NAME                                  READY     STATUS               RESTARTS   AGE
agendor-web-canary-1388324162-2qqgr   1/1       Running              0          5h
agendor-web-canary-1388324162-38xlb   1/1       Running              0          1m
agendor-web-canary-1388324162-3g815   1/1       Running              0          5h
agendor-web-canary-1388324162-89j6k   1/1       Running              0          5h
agendor-web-canary-1388324162-gjndp   0/1       ContainerCreating    0          4s
agendor-web-canary-1388324162-n57wc   0/1       ContainerCreating    0          4s
agendor-web-canary-1388324162-q0pcv   0/1       ContainerCreating    0          4s
agendor-web-canary-1388324162-t8v4s   0/1       ContainerCreating    0          4s

Limits tunning

# Default values for agendor-web.
# This is a YAML-formatted file.
# Declare variables to be passed into your templates.
name: agendor-web
namespace: default
app_env: staging

# autoscale configurations
minReplicas: 2
maxReplicas: 10
targetCPUUtilizationPercentage: 50


...
resources:
  limits:
    cpu: 100m
    memory: 256Mi
  requests:
    cpu: 100m
    memory: 256Mi

# Default values for agendor-web.
# This is a YAML-formatted file.
# Declare variables to be passed into your templates.
name: agendor-web
namespace: default
app_env: staging

# autoscale configurations
minReplicas: 2
maxReplicas: 10
targetCPUUtilizationPercentage: 30 # decreased


...
resources:
  limits:
    cpu: 200m # increased
    memory: 512Mi # increased
  requests:
    cpu: 100m
    memory: 256Mi

Pods Scaling + Deploy

Rolling Update Fail

Deployment vs HPA

kind: Deployment
apiVersion: extensions/v1beta1
metadata:
  name: {{ .Values.name }}
...
spec:
  replicas: {{ .Values.minReplicas }}
  revisionHistoryLimit: 10
  selector:
    matchLabels:
      app: {{ .Values.name }}
apiVersion: autoscaling/v1
kind: HorizontalPodAutoscaler
metadata:
  name: {{ .Values.name }}
...
spec:
  scaleTargetRef:
    apiVersion: extensions/v1beta1
    kind: Deployment
    name: {{ .Values.name }}
  minReplicas: {{ .Values.minReplicas }}
  maxReplicas: {{ .Values.maxReplicas }}
  targetCPUUtilizationPercentage: {{ .Values.targetCPUUtilizationPercentage }}

# values.yaml
minReplicas: 1
maxReplicas: 10
targetCPUUtilizationPercentage: 30

strategy:
  rollingUpdate:
    maxSurge: 3
    maxUnavailable: 0

Rolling Update Win

Canary Bug Tracking

Canary Bug Tracking

Traefik redireciona aleatoriamente

Cliente reclama de bug: problema na aplicação ou na nova infra?

Redirecionamento aleatório:

Bom para testar escalabilidade e performance

Ruim se mantido por muito tempo

Canary Bug Tracking

# aws.yml (traefik)
backends:
  web:
    - name: legacy_1
      values:
        - weight = 5
        - url = "http://ec2-54-94-240-102.sa-east-1.compute.amazonaws.com"
    - name: legacy_2
      values:
        - weight = 5
        - url = "http://ec2-54-233-74-185.sa-east-1.compute.amazonaws.com"
    - name: k8s
      values:
        - weight = 1
        - url = "http://agendor-api-v2-canary.production.svc.cluster.local"
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"
# aws.yml (traefik)
backends:
  web:
    - name: legacy_1
      url = "http://ec2-54-94-240-102.sa-east-1.compute.amazonaws.com"
    - name: legacy_2
      url = "http://ec2-54-233-74-185.sa-east-1.compute.amazonaws.com"
  web_canary:
    - name: k8s
      url: "http://agendor-web-canary.production.svc.cluster.local"
# PHP
public function addBetaTesterCookie($accountId = NULL) {
  if ($this->isCanaryBetaTester($accountId)) { // check on Redis
    $this->setCookie(BetaController::$BETA_COOKIE_NAME, 
        BetaController::$BETA_COOKIE_VALUE);
  } else {
    $this->setCookie(BetaController::$BETA_COOKIE_NAME, 
        BetaController::$LEGACY_COOKIE_VALUE);
  }
}

Cluster Scaling #1

Aviso de lentidão no sistema

Análise da lentidão

Análise da lentidão

SSH into Server

$ ssh -i ~/.ssh/kube_aws_rsa 
    admin@ec2-XX-XXX-XXX-XXX.sa-east-1.compute.amazonaws.com
// Deveria ter usado:
$ kubectl top nodes

Debugging Nodes 

$ kubectl describe nodes

Debugging auto scaler

I0622 18:08:59 scale_up.go:62] Upcoming 1 nodes
I0622 18:08:59 scale_up.go:81] Skipping node group nodes.k8s.agendor.com.br - max size reached
I0622 18:08:59 scale_up.go:132] No expansion options
kubectl logs -f --tail=100 --namespace=kube-system cluster-autoscaler-2462627368-c1hhl
containers:
  - name: cluster-autoscaler
    image: "{{ .Values.image.repository }}:{{ .Values.image.tag }}"
    imagePullPolicy: "{{ .Values.image.pullPolicy }}"
    command:
      - ./cluster-autoscaler
      - --cloud-provider=aws
    {{- range .Values.autoscalingGroups }}
      - --nodes={{ .minSize }}:{{ .maxSize }}:{{ .name }}
    {{- end }}
autoscalingGroups:
  - name: nodes.k8s.agendor.com.br
    minSize: 1
    maxSize: 15

values.yaml

deployment.yaml

Change instance groups

$ kops get ig
Using cluster from kubectl context: k8s.agendor.com.br

NAME			ROLE	MACHINETYPE	MIN	MAX	SUBNETS
master-sa-east-1a	Master	t2.medium	1	1	sa-east-1a
nodes			Node	t2.medium	2	6	sa-east-1a,sa-east-1c
apiVersion: kops/v1alpha2
kind: InstanceGroup
metadata:
  creationTimestamp: "2017-02-04T07:30:23Z"
  labels:
    kops.k8s.io/cluster: k8s.agendor.com.br
  name: nodes
spec:
  image: kope.io/k8s-1.6-debian-jessie-amd64-hvm-ebs-2017-05-02
  machineType: t2.medium
  maxSize: 20
  minSize: 2
  role: Node
  rootVolumeSize: 50
  subnets:
  - sa-east-1a
  - sa-east-1c
$ kops edit ig nodes
apiVersion: kops/v1alpha2
kind: InstanceGroup
metadata:
  creationTimestamp: "2017-02-04T07:30:23Z"
  labels:
    kops.k8s.io/cluster: k8s.agendor.com.br
  name: nodes
spec:
  image: kope.io/k8s-1.6-debian-jessie-amd64-hvm-ebs-2017-05-02
  machineType: t2.medium
  maxSize: 6
  minSize: 2
  role: Node
  rootVolumeSize: 50
  subnets:
  - sa-east-1a
  - sa-east-1c

Change instance groups

$ kops update cluster
Using cluster from kubectl context: k8s.agendor.com.br


*********************************************************************************

A new kubernetes version is available: 1.6.6
Upgrading is recommended (try kops upgrade cluster)

More information: https://github.com/kubernetes/kops/blob/master/permalinks/upgrade_k8s.md#1.6.6

*********************************************************************************

I0623 10:13:07.893227   35173 executor.go:91] Tasks: 0 done / 60 total; 27 can run
I0623 10:13:10.226305   35173 executor.go:91] Tasks: 27 done / 60 total; 14 can run
I0623 10:13:10.719476   35173 executor.go:91] Tasks: 41 done / 60 total; 17 can run
I0623 10:13:11.824240   35173 executor.go:91] Tasks: 58 done / 60 total; 2 can run
I0623 10:13:11.933869   35173 executor.go:91] Tasks: 60 done / 60 total; 0 can run
Will modify resources:
  AutoscalingGroup/nodes.k8s.agendor.com.br
  	MaxSize             	 6 -> 20


Must specify --yes to apply changes

From kops update cluster docs:

This change will apply to new instances only; if you'd like to roll it out immediately to all the instances you have to perform a rolling update.

 

See a preview with: kops rolling-update cluster

Then restart the machines with: kops rolling-update cluster --yes

Cluster Scaling #2

Black Hole

kubectl logs -f --tail=200 --namespace=kube-system \
    $(kubectl get pods --namespace=kube-system \
        | grep cluster-autoscaler \
        | awk '{print $1}' \
        | head -1)
kubectl describe nodes
$ aws ec2 describe-route-tables \
    --filters "Name=route-table-id, Values=rtb-7923f31d" \
    | grep blackhole \
    | awk '{print $2}' \
    | xargs -n 1 \
        aws ec2 delete-route --route-table-id rtb-7923f31d --destination-cidr-block
12:51:26 scale_up.go:62] Upcoming 0 nodes
12:51:26 scale_up.go:145] Best option to resize: nodes.k8s.agendor.com.br
12:51:26 scale_up.go:149] Estimated 1 nodes needed in nodes.k8s.agendor.com.br
12:51:26 scale_up.go:169] Scale-up: setting group nodes.k8s.agendor.com.br size to 6
12:51:26 aws_manager.go:124] Setting asg nodes.k8s.agendor.com.br size to 6
12:51:26 event.go:217] Event(v1.ObjectReference{Kind:"ConfigMap", Namespace:"kube-system", Name:"cluster-autoscaler-status", UID:"3d1fa43e-59e8-11e7-b0ba-02fccc543533", APIVersion:"v1", ResourceVersion:"22154083", FieldPath:""}): type: 'Normal' reason: 'ScaledUpGroup' Scale-up: group nodes.k8s.agendor.com.br size set to 6
12:51:26 event.go:217] Event(v1.ObjectReference{Kind:"Pod", Namespace:"production", Name:"agendor-web-canary-3076992955-95b3p", UID:"2542369b-5a6e-11e7-b0ba-02fccc543533", APIVersion:"v1", ResourceVersion:"22154097", FieldPath:""}): type: 'Normal' reason: 'TriggeredScaleUp' pod triggered scale-up, group: nodes.k8s.agendor.com.br, sizes (current/new): 5/6
...
12:51:37 scale_up.go:62] Upcoming 1 nodes
12:51:37 scale_up.go:124] No need for any nodes in nodes.k8s.agendor.com.br
12:51:37 scale_up.go:132] No expansion options
12:51:37 static_autoscaler.go:247] Scale down status: unneededOnly=true lastScaleUpTime=2017-06-26 12:51:26.777491768 +0000 UTC lastScaleDownFailedTrail=2017-06-25 20:52:45.172500628 +0000 UTC schedulablePodsPresent=false
12:51:37 static_autoscaler.go:250] Calculating unneeded nodes

Deletar rotas em status "Black Hole"

$ kubectl logs -f --tail=100 --namespace=kube-system 
    $(kubectl get pods --namespace=kube-system |
        grep cluster-autoscaler | awk '{print $1}' | head -1)

Lições Aprendidas

Lições Aprendidas

Múltiplas versões do código simultâneas

Rolling Update

Canary Release

Todos devem desenvolver pensando no deploy/release desde o começo

Lições Aprendidas

Muitas tecnologias novas de uma vez

Helm

Traefik

Docker

AWS

Kubernetes

Quando usadas em conjunto, o risco aumenta

Kops

Cluster Auto Scaler

Assim como a curva de aprendizado

Lições Aprendidas

Helm

--recreate-pods para corrigir o problema dos pods não serem re-criados

Impacto: Rolling Update fail

   env:
    - name: APP_NAME
      value: "{{ .Values.name }}"
    - name: APP_NAMESPACE
      value: "{{ .Values.namespace }}"
    - name: APP_ENV
      value: "{{ .Values.app_env }}"
    - name: APP_FULLNAME
      value: "{{ .Values.name }}-{{ .Values.app_env }}"
    # Force recreate pods
    - name: UPDATE_CONFIG
      value: {{ $.Release.Time }}

Obrigado!

Dúvidas?

linkedin.com/in/tuliomonteazul

@tuliomonteazul

tulio@agendor.com.br

Everton Ribeiro

Gullit Miranda

Marcus Gadbem

Special thanks to

Copy of Canary Release com Kubernetes e Traefik

By Tulio Monte Azul

Made with Slides.com

Copy of Canary Release com Kubernetes e Traefik

  • 94

More from Tulio Monte Azul