VCS GitOps

GitOps setup with ArgoCD

This page provides a detailed overview of how to set up the GitOps functionality for Seldon Deploy. Setting up GitOps is highly recommended.


Seldon Deploy leverages ArgoCD to implement GitOps in your cluster. You can read more about Seldon Deploy’s GitOps integration in the architecture and operation sections.

This documentation page will walk you through the different steps required to set up ArgoCD and GitOps in your cluster. Note that this guide assumes a fresh cluster without ArgoCD, thus it must only be treated as a general reference. Feel free to adapt these steps to your particular infrastructure.

Installation of ArgoCD

Since this guide assumes that you currently don’t have ArgoCD installed in your cluster, the first step will be on installing and configuring it. You can treat these steps as a quick way to get started with ArgoCD. Therefore, it is highly encouraged to check ArgoCD’s official documentation for further configuration.

To install and configure ArgoCD we will first install their official K8s manifest into our cluster. For version v1.6.2 of ArgoCD, this can be done as follows:

ARGOCD_VERSION=v1.6.2

kubectl create namespace argocd || echo "namespace exist"
kubectl apply -n argocd -f https://raw.githubusercontent.com/argoproj/argo-cd/${ARGOCD_VERSION}/manifests/install.yaml

kubectl rollout status deployment/argocd-application-controller -n argocd
kubectl rollout status deployment/argocd-repo-server -n argocd
kubectl rollout status deployment/argocd-server -n argocd
kubectl rollout status deployment/argocd-redis -n argocd
kubectl rollout status deployment/argocd-dex-server -n argocd

Configuration of ArgoCD

Now that we have ArgoCD installed in our cluster, the next step will involve configuring it so that:

  • The admin user credentials get changed.
  • SSL termination is disabled in ArgoCD.
  • ArgoCD gets exposed under the /argocd/ path of our load balancer.

Note that these steps must be treated just as a guideline. Therefore, you may need to adapt them to your infrastructure. For more information on ArgoCD’s configuration, we suggest checking ArgoCD’s official documentation.

Admin user and password

Upon the initial installation of ArgoCD, it is recommended to change its user and password. This guideline will show how to change ArgoCD’s default admin user as an example. However, it’s also possible to hook ArgoCD to a centralised OIDC provider. This could be the same as Seldon Deploy’s OIDC provider, providing a Single Sign-On across your entire stack.

The following command will assume that we want to set the admin credentials as admin // $ARGOCDINITIALPASS. With this in mind, we can set those credentials as:

ARGOCDINITIALPASS=12341234
kubectl patch secret \
    -n argocd argocd-secret \
    -p '{"stringData": { "admin.password": "'$(htpasswd -bnBC 10 "" ${ARGOCDINITIALPASS} | tr -d ':\n')'"}}'

This command requires htpasswd that can be installed on Ubuntu, Debian and Mint systemms with:

sudo apt install apache2-utils

SSL Termination

As a general choice, we recommend to handle SSL termination at the ingress / load balancer level. This simplifies the components setup within the cluster.

Following this approach, we will need to disable the SSL termination in ArgoCD. Otherwise, ArgoCD will expect to receive SSL traffic by default. To do this, we can ask ArgoCD to run in insecure mode by running the command below:

kubectl patch deploy argocd-server \
    -n argocd \
    -p '[{"op": "add", "path": "/spec/template/spec/containers/0/command/-", "value": "--insecure"}]' \
    --type json

Exposing ArgoCD through our ingress

There are different ways to access ArgoCD. However, the most convenient is usually to expose it through our ingress layer so that it can be accessed as a different subpath (e.g. /argocd/). This will make the ArgoCD API and UI reachable as http(s)://<load-balancer-domain>/argocd/.

This will need to be configured in a couple of places.

  1. On one hand, we need to tell ArgoCD to expect its static assets to be served under the /argocd/ path. This can be done by running the following command:

    kubectl patch deploy argocd-server \
       -n argocd \
       -p '[{"op": "add", "path": "/spec/template/spec/containers/0/command/-", "value": "--rootpath"}, {"op": "add", "path": "/spec/template/spec/containers/0/command/-", "value": "/argocd"}]' \
       --type json
  2. On the other hand, we will need to create a VirtualService that maps the /argocd/ path to ArgoCD’s service. Note that this assumes that we are using Istio as an ingress layer. To expose ArgoCD, it should be enough to apply the following VirtualService resource to our cluster:

    cat << EOF > ./argocd-vs.yaml
    apiVersion: networking.istio.io/v1alpha3
    kind: VirtualService
    metadata:
     name: argocd-server
     namespace: argocd
    spec:
     gateways:
     - istio-system/seldon-gateway
     hosts:
     - '*'
     http:
     - match:
       - uri:
           prefix: /argocd/
       route:
       - destination:
           host: argocd-server
           port:
             number: 80
    EOF

If we assume that the above has been saved as a file called argocd-vs.yaml, we can then apply it as:

   kubectl apply -f argocd-vs.yaml

Accessing ArgoCD UI and API

After configuring ArgoCD, we should now be able to reach ArgoCD under:

http(s)://<load-balancer-domain>/argocd/

In a standard setup, you should be able to obtain your load balancer domain by checking the Istio resources created in your cluster. You can do this as:

ISTIO_INGRESS=$(\
  kubectl get svc -n istio-system istio-ingressgateway \
    -o jsonpath='{.status.loadBalancer.ingress[0].ip}'\
)
ISTIO_INGRESS+=$(\
  kubectl get svc -n istio-system istio-ingressgateway \
  -o jsonpath='{.status.loadBalancer.ingress[0].hostname}'\
)

echo $ISTIO_INGRESS

Alternatively, if your ingress layer doesn’t expose an external IP address or domain, it’s also possible to port-forward a local port into ArgoCD. For example, to map your local localhost:8080 (or localhost:8080/argocd/ if you defined the rootpath above) domain to ArgoCD, you can run:

kubectl port-forward -n argocd svc/argocd-server 8080:80

Once we know our load balancer domain, we can then access ArgoCD through its UI or API. For the latter, we will leverage ArgoCD’s CLI, called argocd. Below, you can find instructions on how to set up and access both of them.

ArgoCD ships with a CLI tool called argocd. This tool allows you to interact with ArgoCD from the command line. The rest of this guide will provide command examples that can be used to configure ArgoCD using their argocd tool. However, it’s also possible to perform the same actions from the ArgoCD UI.

To use the argocd CLI, we will need to:

  1. Install the argocd CLI locally. This can be achieved by running the following commands:

    ARGOCD_VERSION=v1.6.2
    wget -q -O argocd "https://github.com/argoproj/argo-cd/releases/download/${ARGOCD_VERSION}/argocd-linux-amd64"
    chmod +x ./argocd
  2. Authenticate the CLI against our ArgoCD instance. Assuming that our load balancer domain can be found under $ISTIO_INGRESS, it should be enough to run the command below.

    ./argocd login \
     "$ISTIO_INGRESS:80" \
     --username admin \
     --password ${ARGOCDINITIALPASS} \
     --insecure \
     --grpc-web-root-path /argocd

ArgoCD exposes a UI that can be used to configure your GitOps repository. Assuming that you have followed the ArgoCD setup guidelines above, this UI should be accessible under the /argocd/ path of your load balancer domain.

After accessing the UI, you should be prompted with a login form asking for the admin credentials. You should be able to access this UI by using ArgoCD’s user / password combination (admin and $ARGOCDINITIALPASS by default, respectively).

You can learn more about how to navigate ArgoCD’s UI in their documentation.

GitOps for logging components

The request logger is an infrastructure component that runs in the centralised logging namespace (named as seldon-logs by default). Alongside this, Triggers can also be created for detectors and these tied to particular models. Thus is good to back them through a GitOps environment.

With this goal in mind, it’s encouraged to create a system-level GitOps environment to track these infrastructure resources. To do this, you can follow the instructions to create a new GitOps environment, using the seldon-logs namespace as target. Note that if you’ve configured Seldon Deploy to use a different centralised namespace for logging, you’ll need to point to that namespace instead.

Once the logging GitOps environment is created, Seldon Deploy should be able to pick it up and use it without any further configuration.

Creating a new GitOps environment

Now that ArgoCD has been installed and configured, we can then proceed to create a new GitOps environment. This environment will provide a Kubernetes namespace to deploy our machine learning models, which is backed by a GitOps repository.

This section will walk you through how to create a new GitOps-tracked environment for a namespace with name $namespace. The GitOps repository which will act as the source of truth for this environment will exist under $GIT_REPO_URL. Note that $GIT_REPO_URL should not contain the http or https protocol schema.

Git Repository

The only requirements for the Git repository are:

  • It contains a non-empty folder named as the namespace (i.e. ./$namespace).
  • It is accessible from within the cluster. This can be verified using this one-liner command:

    kubectl run --quiet=true -it --rm git-clone --image=radial/busyboxplus:git  --restart=Never -- git clone https://$GIT_USER:$GIT_TOKEN@$GIT_REPO_URL

ArgoCD Resources

In order for ArgoCD to keep track of our new environment, it’s necessary to configure a Project and an Application. These are concepts native to ArgoCD. We suggest checking the ArgoCD official documentation for more information about them.

ArgoCD projects and applications can be created via the ArgoCD UI, argocd CLI tool and through CRDs. As declarative approach guarantees best reproducibility this is the approach that we recommend.

We create ArgoCD project by creating the following AppProject resources named argocd-project.yaml

cat << EOF > ./argocd-project.yaml
apiVersion: argoproj.io/v1alpha1
kind: AppProject
metadata:
  name: seldon
  namespace: argocd
spec:
  description: Seldon Deploy Project
  sourceRepos:
  - https://${GIT_REPO_URL}
  destinations:
  - namespace: ${namespace}
    server: https://kubernetes.default.svc
  clusterResourceWhitelist:
  - group: '*'
    kind: '*'
  roles:
  - name: seldon-admin
    policies:
    - p, proj:seldon:seldon-admin, applications, get, seldon/*, allow
    - p, proj:seldon:seldon-admin, applications, create, seldon/*, allow
    - p, proj:seldon:seldon-admin, applications, update, seldon/*, allow
    - p, proj:seldon:seldon-admin, applications, delete, seldon/*, allow
    - p, proj:seldon:seldon-admin, applications, sync, seldon/*, allow
EOF

kubectl apply -f argocd-project.yaml

Similarly we create the ArgoCD Application resource:

ARGO_APP_NAME=seldon-gitops-"${namespace}"

cat << EOF > ./argocd-application.yaml
apiVersion: argoproj.io/v1alpha1
kind: Application
metadata:
  name: ${ARGO_APP_NAME}
  namespace: argocd
spec:
  project: seldon
  destination:
    namespace: ${namespace}
    server: https://kubernetes.default.svc
  source:
    directory:
      recurse: true
    path: ${namespace}
    repoURL: https://${GIT_REPO_URL}
  syncPolicy:
    automated: {}
EOF

kubectl apply -f argocd-application.yaml

Kubernetes Model Namespace

To use a namespace with GitOps, you will need to annotate it with the Git repository URL (using key git-repo) and enable it for GitOps access by adding the git-repo label. As an example, to enable a namespace named $namespace, you could run the following commands:

kubectl create ns $namespace || echo "namespace $namespace already exists"

kubectl label ns $namespace seldon.restricted=false --overwrite=true
kubectl label namespace $namespace seldon.gitops=enabled --overwrite=true
kubectl annotate namespace $namespace git-repo="https://${GIT_REPO_URL}" --overwrite=true

If you intend to use Batch jobs on the namespace then you’ll need additional config. You can visit the Argo section of the docs for more details.

Configuring Git Credentials

The last required step is to provide the relevant Git credentials to access our GitOps repository. Note that these credentials need to be set at both the ArgoCD and Seldon Deploy level.

ArgoCD

You can configure your GitOps repository access in ArgoCD using either SSH or user / password credentials. These credentials can be provided using the argocd CLI tool, or directly through the ArgoCD UI.

Assuming that the argocd CLI tool is already authenticated, and that $GIT_SSH_PATH and $GIT_TOKEN represent our GitOps repository user and password, we should be able to run:

./argocd repo add "ssh://${GIT_REPO_URL}" \
    --ssh-private-key-path ${GIT_SSH_PATH} \
    --upsert

Assuming that the argocd CLI tool has already been authenticated, and that ${GIT_USER} and ${GIT_TOKEN} represent our GitOps repository user and password, we should be able to run:

./argocd repo add "https://${GIT_REPO_URL}" \
    --username ${GIT_USER} \
    --password ${GIT_TOKEN} \
    --upsert

ArgoCD exposes a UI that can be used to configure your GitOps repository. Assuming that you have followed the ArgoCD setup guidelines above, this UI should be accessible under the /argocd path. You should be able to access this UI by using ArgoCD’s user / password combination (admin and 12341234 by default, respectively).

You can learn more about how to configure repositories using ArgoCD’s UI in their documentation.

Seldon Deploy

Seldon Deploy also requires access to the repository. This access will be used to monitor the status of the current resources in the cluster, as well as to create new ones.

To configure our Git credentials in Seldon Deploy, we will follow these steps:

  1. Create a Kubernetes secret containing our credentials, either as a SSH key or a User / Password combination. This secret can have any arbitrary name, but must live in the same namespace as Seldon Deploy.

If we assume that our private key is present under $GIT_SSH_PATH, we can create the credentials secret as:

   kubectl create secret generic git-creds -n seldon-system \
     --from-file ${GIT_SSH_PATH} \
     --from-file ${GIT_KNOWN_HOSTS_PATH} \
     --from-literal=passphrase="${GIT_SSHKEY_PASSPHRASE}" \
     --from-literal=username="${GIT_USER}" \
     --from-literal=email="${GIT_EMAIL}"  \
     --dry-run=client -o yaml | kubectl apply -f -

Note that the passphrase field can be left empty if they SSH key doesn’t have a passphrase.

We can create the credentials secret using a User / Password combination (or User / Personal Access Token) as:

   kubectl create secret generic git-creds -n seldon-system \
     --from-literal=username="${GIT_USER}" \
     --from-literal=token="${GIT_TOKEN}" \
     --from-literal=email="${GIT_EMAIL}"  \
     --dry-run=client -o yaml | kubectl apply -f -
  1. Update Seldon Deploy’s configuration to point to our newly created secret. In particular, we will need to modify the gitops section of the values of the Seldon Deploy Helm chart. Here, we will need to set the gitops.argocd.enabled flag to true, and we will point the gitops.git.secret field to the right secret name (e.g. git-creds if we’ve followed the commands above).

    gitops:
     git:
       secret: git-creds
     argocd:
       enabled: true

(Optional) Webhooks Configuration

By default, the resource synchronisation against the GitOps repository will happen on a poll basis. That is, Seldon Deploy will check the repo periodically for updates.

The main caveat of this approach is that there may be a small delay between taking an action and seeing it reflected in the cluster / UI. To work around this, you can configure a set of webhooks in the cluster in the that will get pinged everytime there is a change in the cluster. Note that this is an optional step, and its feasibility may depend on your networking infrastructure and your Git provider.

You can find instructions below on how to set up these webhooks for common Git providers

First we need to configure the webhook that will be sending updates to the Seldon Deploy endpoint:

whIp=$(kubectl get service seldon-deploy-webhook -o jsonpath='{.status.loadBalancer.ingress[*].ip}')
curl -u ${GIT_USER}:${GIT_TOKEN} \
     -v -H "Content-Type: application/json" \
     -d '
      {
        "name": "web", 
        "active": true, 
        "events": ["*"],
        "config": {
          "url": "http://${whIp}/api/git-webhook",
          "content_type": "json"
        }
      }' \
     https://${GIT_API}/repos/${GIT_USER}/${REPONAME}/hooks

Then we set up the webhook that will be sending the updates to ArgoCD.

curl -u ${GIT_USER}:${GIT_TOKEN} \
     -v -H "Content-Type: application/json" \
     -d '\
      {
        "name": "web", 
        "active": true, 
        "events": ["*"],
        "config": {
          "url": "https://${ARGOCDURL}/api/webhook", 
          "content_type": "json",
          "secret": "${ARGOCDINITIALPASS}",
          "insecure_ssl": "1"
        }
      }' \
     https://${GIT_API}/repos/${GIT_USER}/${REPONAME}/hooks

First we need to configure the webhook that will be sending updates to the Seldon Deploy endpoint:

whIp=$(kubectl get service seldon-deploy-webhook -o jsonpath='{.status.loadBalancer.ingress[*].ip}')
curl -u ${GIT_USER}:${GIT_TOKEN} \
     -v -H "Content-Type: application/json" \
     -d '
      {
          "description": "web",
          "events": ["repo:push"],
          "url": "'"http://${whIp}/api/git-webhook"'",
          "active": true
      }' \
    https://${GIT_API}/2.0/repositories/${GIT_USER}/${REPONAME}/hooks

Then we set up the webhook that will be sending the updates to ArgoCD.

curl -u ${GIT_USER}:${GIT_TOKEN} \
     -v -H "Content-Type: application/json" \
     -d '
      {
          "description": "web",
          "events": ["repo:push"],
          "url": "'"https://${ARGOCDURL}/api/webhook"'",
          "active": true
      }' \
    https://${GIT_API}/2.0/repositories/${GIT_USER}/${REPONAME}/hooks

Troubleshooting

You should see one application per gitops namespace when running the below:

kubectl get application -n argocd

You can check the status of an argocd app with kubectl get application -n argocd seldon-gitops-<namespace_name> -o yaml.

The argocd UI (described above) can also be used to inspect applications. You can trigger a manual sync from there. If a particular application is missing or in error try running its setup again in order to debug.

Example Script for Namespace Setup

Download installation resources as explained here

Copy the sd-setup directory containing gitops-setup.sh

cp -r ./seldon-deploy-install/sd-setup/ sd-setup/

The gitops-setup.sh is an example script for setting up gitops in the cluster by installing argocd and configuring namespaces. It can be used for reference.

See the trial installation docs for more on how to run that script. More context is provided in Namespace Setup

Last modified April 20, 2021