Machine learning with Vineyard on Kubernetes#
In this demonstration, we will build a fraudulent transaction classifier for fraudulent transaction data. The process consists of the following three main steps:
prepare-data: Utilize Vineyard to read and store data in a distributed manner.process-data: Employ Mars to process the data across multiple nodes.train-data: Use Pytorch to train the model on the distributed data.
We have three tables: user table, product table, and transaction table.
The user and product tables primarily contain user and product IDs, along with
their respective Feature vectors. Each record in the transaction table indicates
a user purchasing a product, with a Fraud label identifying whether the
transaction is fraudulent. Additional features related to these transactions are also
stored in the transaction table. You can find the three tables in the dataset repo.
Follow the steps below to reproduce the demonstration. First, create a vineyard cluster
with 3 worker nodes.
$ cd k8s && make -C k8s/test/e2e install-vineyard-cluster
Expected output
the kubeconfig path is /tmp/e2e-k8s.config Creating the kind cluster with local registry a16c878c5091c1e5c9eff0a1fca065665f47edb4c8c75408b3d33e22f0ec0d05 Creating cluster "kind" ... ✓ Ensuring node image (kindest/node:v1.24.0) 🖼 ✓ Preparing nodes 📦 📦 📦 📦 ✓ Writing configuration 📜 ✓ Starting control-plane 🕹️ ✓ Installing CNI 🔌 ✓ Installing StorageClass 💾 ✓ Joining worker nodes 🚜 Set kubectl context to "kind-kind" You can now use your cluster with: kubectl cluster-info --context kind-kind --kubeconfig /tmp/e2e-k8s.config Thanks for using kind! 😊 configmap/local-registry-hosting created Installing vineyard-operator... The push refers to repository [localhost:5001/vineyard-operator] c3a672704524: Pushed b14a7037d2e7: Pushed 8d7366c22fd8: Pushed latest: digest: sha256:ea06c833351f19c5db28163406c55e2108676c27fdafea7652500c55ce333b9d size: 946 make[1]: Entering directory '/opt/caoye/v6d/k8s' go: creating new go.mod: module tmp /home/gsbot/go/bin/controller-gen rbac:roleName=manager-role crd:maxDescLen=0 webhook paths="./..." output:crd:artifacts:config=config/crd/bases cd config/manager && /usr/local/bin/kustomize edit set image controller=localhost:5001/vineyard-operator:latest /usr/local/bin/kustomize build config/default | kubectl apply -f - namespace/vineyard-system created customresourcedefinition.apiextensions.k8s.io/backups.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/globalobjects.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/localobjects.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/operations.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/recovers.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/sidecars.k8s.v6d.io created customresourcedefinition.apiextensions.k8s.io/vineyardds.k8s.v6d.io created serviceaccount/vineyard-manager created role.rbac.authorization.k8s.io/vineyard-leader-election-role created clusterrole.rbac.authorization.k8s.io/vineyard-manager-role created clusterrole.rbac.authorization.k8s.io/vineyard-metrics-reader created clusterrole.rbac.authorization.k8s.io/vineyard-proxy-role created clusterrole.rbac.authorization.k8s.io/vineyard-scheduler-plugin-role created rolebinding.rbac.authorization.k8s.io/vineyard-leader-election-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-kube-scheduler-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-manager-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-proxy-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-scheduler-plugin-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-scheduler-rolebinding created clusterrolebinding.rbac.authorization.k8s.io/vineyard-volume-scheduler-rolebinding created service/vineyard-controller-manager-metrics-service created service/vineyard-webhook-service created deployment.apps/vineyard-controller-manager created mutatingwebhookconfiguration.admissionregistration.k8s.io/vineyard-mutating-webhook-configuration created validatingwebhookconfiguration.admissionregistration.k8s.io/vineyard-validating-webhook-configuration created make[1]: Leaving directory '/opt/caoye/v6d/k8s' deployment.apps/vineyard-controller-manager condition met Vineyard-Operator Ready Installing vineyard cluster... vineyardd.k8s.v6d.io/vineyardd-sample created vineyardd.k8s.v6d.io/vineyardd-sample condition met Vineyard cluster Ready
Verify that all Vineyard pods are running.
$ KUBECONFIG=/tmp/e2e-k8s.config kubectl get pod -n vineyard-system
Expected output
NAME READY STATUS RESTARTS AGE etcd0 1/1 Running 0 68s etcd1 1/1 Running 0 68s etcd2 1/1 Running 0 68s vineyard-controller-manager-7f569b57c5-46tgq 2/2 Running 0 92s vineyardd-sample-6ffcb96cbc-gs2v9 1/1 Running 0 67s vineyardd-sample-6ffcb96cbc-n59gg 1/1 Running 0 67s vineyardd-sample-6ffcb96cbc-xwpzd 1/1 Running 0 67s
First, let’s prepare the dataset and download it into the kind worker nodes as follows.
$ worker=($(docker ps | grep kind-worker | awk -F ' ' '{print $1}'))
$ for c in ${worker[@]}; do \
docker exec $c sh -c "\
mkdir -p /datasets; \
cd /datasets/; \
curl -OL https://raw.githubusercontent.com/GraphScope/gstest/master/vineyard-mars-showcase-dataset/{item,txn,user}.csv" \
done
The prepare-data job primarily reads the datasets and distributes them across different Vineyard nodes. For more information, please refer to the prepare data code. To apply the job, follow the steps below:
Note
The prepare-data job needs to exec into the other pods. Therefore, you need to create a service account and bind it to the role under the namespace. Please make sure you can have permission to create the following role.
- apiGroups: [""]
resources: ["pods", "pods/log", "pods/exec"]
verbs: ["get", "patch", "delete", "create", "watch", "list"]
$ kubectl create ns vineyard-job && \
kubectl apply -f showcase/vineyard-mars-pytorch/prepare-data/resources && \
kubectl wait job -n vineyard-job -l app=prepare-data --for condition=complete --timeout=1200s
Expected output
namespace/vineyard-job created clusterrolebinding.rbac.authorization.k8s.io/prepare-data-rolebinding created clusterrole.rbac.authorization.k8s.io/prepare-data-role created job.batch/prepare-data created serviceaccount/prepare-data created job.batch/prepare-data condition met
Note
The process-data job needs to create a new namespace and deploy several kubernetes resources in it. Please make sure you can have permission to create the following role.
- apiGroups: [""]
resources: ["pods", "pods/exec", "pods/log", "endpoints", "services"]
verbs: ["get", "patch", "delete", "create", "watch", "list"]
- apiGroups: [""]
resources: ["namespaces"]
verbs: ["get", "create", "delete"]
- apiGroups: [""]
resources: ["nodes"]
verbs: ["get", "list"]
- apiGroups: ["rbac.authorization.k8s.io"]
resources: ["roles", "rolebindings"]
verbs: ["patch", "get", "create", "delete"]
- apiGroups: ["apps"]
resources: ["deployments"]
verbs: ["create"]
Notice, the process-data job will require lots of permissions to deal kubernetes resources, so please check the image of process-data job if it is an official one.
The prepare-data job creates numerous dataframes in Vineyard. To combine these dataframes, we use the appropriate join method in mars. For more details, refer to the process data code. Apply the process-data job as follows:
$ kubectl apply -f showcase/vineyard-mars-pytorch/process-data/resources && \
kubectl wait job -n vineyard-job -l app=process-data --for condition=complete --timeout=1200s
Finally, apply the train-data job to obtain the fraudulent transaction classifier. You can also view the train data code.
$ kubectl apply -f k8s/showcase/vineyard-mars-pytorch/train-data/resources && \
kubectl wait pods -n vineyard-job -l app=train-data --for condition=Ready --timeout=1200s
If any of the above steps fail, please refer to the mars showcase e2e test for further guidance.