Etsuji Nakai | Solutions Architect | Google

Contributed by Google employees.

This tutorial explains how you can use Workflows and other serverless services, such as Cloud Run, to create a custom machine learning (ML) pipeline. The machine learning use case is based on the babyweight model example.

The following diagram shows the overall architecture of what you build in this tutorial:

In this tutorial, you deploy two microservices on Cloud Run: One microservice is to launch a Dataflow pipeline to preprocess the training data. The original data stored in BigQuery is converted to CSV files and stored in a Cloud Storage bucket. The other microservice is to launch a machine learning training job on AI Platform and deploy the trained model for predictions. The machine learning model files are cloned from the GitHub repository.

You deploy a Workflows template to automate the whole process.

• Deploy a microservice that launchs a Dataflow pipeline.
• Deploy a microservice that launchs a machine learning training job on AI Platform and deploy the trained model for predictions.
• Test the deployed microservices using the curl command.
• Deploy a Workflows template to automate the whole process.
• Execute a Workflows job.

This tutorial uses billable components of Google Cloud, including the following:

• Workflows
• Cloud Run
• Dataflow
• AI Platform
• Cloud Build
• Cloud Storage

Use the pricing calculator to generate a cost estimate based on your projected usage.

1. Create a project in the Cloud Console.

2. Enable billing for your project.

3. Open Cloud Shell.

4. Set environment variables for your project ID, the GitHub repository URL, and the directory path to the machine learning model:

   PROJECT_ID="[YOUR_PROJECT_ID]"
   GIT_REPO="https://github.com/GoogleCloudPlatform/community"
   MODEL_PATH="tutorials/ml-pipeline-with-workflows/babyweight_model"
   

   Replace [YOUR_PROJECT_ID] with your project ID.

5. Set the project ID for the Cloud SDK:

   gcloud config set project $PROJECT_ID
   

6. Enable APIs:

   gcloud services enable run.googleapis.com
   gcloud services enable workflows.googleapis.com
   gcloud services enable cloudbuild.googleapis.com
   gcloud services enable dataflow.googleapis.com
   gcloud services enable ml.googleapis.com
   gcloud services enable bigquery.googleapis.com
   

7. Set the storage bucket name in an environment variable and create the bucket:

   BUCKET=gs://$PROJECT_ID-pipeline
   gsutil mb $BUCKET
   

8. Clone the repository:

   cd $HOME
   git clone $GIT_REPO
   

1. Deploy a microservice to preprocess the training data:

   cd $HOME/community/tutorials/ml-pipeline-with-workflows/services/preprocess
   gcloud builds submit --tag gcr.io/$PROJECT_ID/preprocess-service
   gcloud run deploy preprocess-service \
     --image gcr.io/$PROJECT_ID/preprocess-service \
     --platform=managed --region=us-central1 \
     --no-allow-unauthenticated \
     --memory 512Mi \
     --set-env-vars "PROJECT_ID=$PROJECT_ID"
   

2. Deploy a microservice to train and deploy the machine learning model:

   cd $HOME/community/tutorials/ml-pipeline-with-workflows/services/train
   gcloud builds submit --tag gcr.io/$PROJECT_ID/train-service
   gcloud run deploy train-service \
     --image gcr.io/$PROJECT_ID/train-service \
     --platform=managed --region=us-central1 \
     --no-allow-unauthenticated \
     --memory 512Mi \
     --set-env-vars "PROJECT_ID=$PROJECT_ID,GIT_REPO=$GIT_REPO,BRANCH='master',MODEL_PATH=$MODEL_PATH"
   

3. Set service URLs in environment variables:

    SERVICE_NAME="preprocess-service"
    PREPROCESS_SERVICE_URL=$(gcloud run services list --platform managed \
        --format="table[no-heading](URL)" --filter="SERVICE:${SERVICE_NAME}")
   
    SERVICE_NAME="train-service"
    TRAIN_SERVICE_URL=$(gcloud run services list --platform managed \
        --format="table[no-heading](URL)" --filter="SERVICE:${SERVICE_NAME}")
   

Before automating the whole process with Workflows, you test the microservices using the curl command.

The following command send an API request to launch a Dataflow pipeline to preprocess the training data. The option limit specifies the number of rows to extract from BigQuery. You specify a small number (1,000) for a testing purpose in this example.

curl -X POST -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
  -H "Content-Type: application/json" \
  -d "{\"limit\":\"1000\", \"outputDir\":\"$BUCKET/preproc\"}" \
  -s $PREPROCESS_SERVICE_URL/api/v1/preprocess | jq .

The output looks like the following:

{
  "jobId": "2020-12-13_23_57_52-4099585880410245609",
  "jobName": "preprocess-babyweight-054aeefe-16d2-4a26-a5c2-611a5ece1583",
  "outputDir": "gs://workflows-ml-pipeline-pipeline/preproc/054aeefe-16d2-4a26-a5c2-611a5ece1583"
}

The jobId shows the Job ID of the Dataflow pipeline job, and the preprocessed data is stored under outputDir. Copy the Job ID in jobID and the storage path in outputDir to set them in environment variables:

OUTPUT_DIR="gs://workflows-ml-pipeline-pipeline/preproc/054aeefe-16d2-4a26-a5c2-611a5ece1583"
JOB_ID="2020-12-13_23_57_52-4099585880410245609"

The following command sends an API request to show the job status:

curl -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
  -s $PREPROCESS_SERVICE_URL/api/v1/job/$JOB_ID | jq .

The output looks like the following:

{
  "createTime": "2020-12-14T07:57:54.086857Z",
  "currentState": "JOB_STATE_RUNNING",
  "currentStateTime": "2020-12-14T07:57:59.416039Z",
...
}

Wait about 5 minutes until the currentState becomes JOB_STATE_DONE. You can also check the job status in the Cloud Console.

The following command sends an API request to start an AI Platform job to train the machine learning model. The options numTrainExamples, numEvalExamples, and numEvals specify the numbers of training examples, evaluation examples, and evaluations during the training, respectively. You specify small numbers for a testing purpose in this example.

curl -X POST -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
  -H "Content-Type: application/json" \
  -d "{\"jobDir\": \"$BUCKET/trained_model\", \"dataDir\": \"$OUTPUT_DIR\", \"numTrainExamples\": 5000, \"numEvalExamples\": 1000, \"numEvals\": 2}" \
  -s $TRAIN_SERVICE_URL/api/v1/train | jq .

The output looks like the following:

{
  "createTime": "2020-12-14T08:24:12Z",
  "etag": "zKM8N6bPpVk=",
  "jobId": "train_babyweight_e281aab4_5b4f_40cd_8fe3_f8290037b5fc",
  "state": "QUEUED",
...
    "jobDir": "gs://workflows-ml-pipeline-pipeline/trained_model/e281aab4-5b4f-40cd-8fe3-f8290037b5fc",
    "packageUris": [
      "gs://workflows-ml-pipeline-pipeline/trained_model/e281aab4-5b4f-40cd-8fe3-f8290037b5fc/trainer-0.0.0.tar.gz"
    ],
    "pythonModule": "trainer.task",
    "pythonVersion": "3.7",
    "region": "us-central1",
    "runtimeVersion": "2.2",
    "scaleTier": "BASIC_GPU"
  },
  "trainingOutput": {}
}

The jobId shows the Job ID of the training job, and the trained model is stored under jobDir. Copy the Job ID in jobID and the storage path in jobDir to set them in environment variables:

JOB_ID="train_babyweight_e281aab4_5b4f_40cd_8fe3_f8290037b5fc"
JOB_DIR="gs://workflows-ml-pipeline-pipeline/trained_model/e281aab4-5b4f-40cd-8fe3-f8290037b5fc"

The following command sends an API request to show the job status:

curl -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
  -s $TRAIN_SERVICE_URL/api/v1/job/$JOB_ID | jq .

The output looks like the following:

{
  "createTime": "2020-12-14T08:24:12Z",
  "etag": "rW+uQQbA6bM=",
  "jobId": "train_babyweight_e281aab4_5b4f_40cd_8fe3_f8290037b5fc",
  "state": "PREPARING",
...
}

Wait about 10 minutes until state becomes SUCCEEDED. You can also check the job status in the Cloud Console.

The following command sends an API request to launch an AI Platform job to train the machine learning model:

curl -X POST -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
  -H "Content-Type: application/json" \
  -d "{\"modelName\": \"babyweight_model\", \"versionName\": \"v1\", \"deploymentUri\": \"$JOB_DIR/export\"}" \
  -s $TRAIN_SERVICE_URL/api/v1/deploy | jq .

The options modelName and versionName specify the model name and the version name, respectively.

The output looks like the following:

{
  "metadata": {
    "@type": "type.googleapis.com/google.cloud.ml.v1.OperationMetadata",
    "createTime": "2020-12-14T08:34:36Z",
    "modelName": "projects/workflows-ml-pipeline/models/babyweight_model",
    "operationType": "CREATE_VERSION",
    "version": {
      "createTime": "2020-12-14T08:34:35Z",
      "deploymentUri": "gs://workflows-ml-pipeline-pipeline/trained_model/e281aab4-5b4f-40cd-8fe3-f8290037b5fc/export",
      "etag": "BlXqEgx9VQg=",
      "framework": "TENSORFLOW",
      "machineType": "mls1-c1-m2",
      "name": "projects/workflows-ml-pipeline/models/babyweight_model/versions/v1",
      "pythonVersion": "3.7",
      "runtimeVersion": "2.2"
    }
  },
  "name": "projects/workflows-ml-pipeline/operations/create_babyweight_model_v1-1607934875576"
}

Wait a few minutes for the deployed version to become ready, and run the following command to confirm that the model has been deployed:

gcloud ai-platform models list --region global

The output looks like the following:

Using endpoint [https://ml.googleapis.com/]
NAME              DEFAULT_VERSION_NAME
babyweight_model  v1

You can also check the deployed model in the Cloud Console.

You use Workflows to automate the steps that you've done in the previous section.

Run the following commands to create a service account and assign the role to invoke services on Cloud Run:

SERVICE_ACCOUNT_NAME="cloud-run-invoker"
SERVICE_ACCOUNT_EMAIL=${SERVICE_ACCOUNT_NAME}@${PROJECT_ID}.iam.gserviceaccount.com
gcloud iam service-accounts create $SERVICE_ACCOUNT_NAME \
  --display-name "Cloud Run Invoker"
gcloud projects add-iam-policy-binding $PROJECT_ID \
  --member=serviceAccount:$SERVICE_ACCOUNT_EMAIL \
  --role=roles/run.invoker

Run the following commands to deploy the workflow template, which you associate with the service account that you created in the previous step using the --service-account option:

cd $HOME/community/tutorials/ml-pipeline-with-workflows/workflows
cp ml_workflow.yaml.template ml_workflow.yaml
sed -i "s#PREPROCESS-SERVICE-URL#${PREPROCESS_SERVICE_URL}#" ml_workflow.yaml
sed -i "s#TRAIN-SERVICE-URL#${TRAIN_SERVICE_URL}#" ml_workflow.yaml
gcloud beta workflows deploy ml_workflow \
  --source=ml_workflow.yaml \
  --service-account=$SERVICE_ACCOUNT_EMAIL

Wait a few minutes for the service account to become ready, and then proceed to the next step.

Run the following command to execute a Workflows job:

gcloud beta workflows execute ml_workflow \
  --data="{\"limit\": 1000, \"bucket\": \"$BUCKET\", \"numTrainExamples\": 5000, \"numEvals\": 2, \"numEvalExamples\": 1000, \"modelName\": \"babyweight_model\", \"versionName\": \"v2\"}"

You can monitor the status of the job in the Cloud Console.

When the job has successfully completed, run the following command to confirm that the model has been deployed:

gcloud ai-platform versions list --model babyweight_model --region global

The output looks like the following:

Using endpoint [https://ml.googleapis.com/]
NAME  DEPLOYMENT_URI                                                                                 STATE
v1    gs://workflows-ml-pipeline-pipeline/trained_model/e281aab4-5b4f-40cd-8fe3-f8290037b5fc/export  READY
v2    gs://workflows-ml-pipeline-pipeline/trained_model/9550e422-a3ee-437b-83db-f321610344f3/export  READY

To avoid incurring charges to your Google Cloud account for the resources used in this tutorial, you can delete the project.

Deleting a project has the following consequences:

• If you used an existing project, you'll also delete any other work that you've done in the project.
• You can't reuse the project ID of a deleted project. If you created a custom project ID that you plan to use in the future, delete the resources inside the project instead. This ensures that URLs that use the project ID, such as an appspot.com URL, remain available.

To delete a project, do the following:

1. In the Cloud Console, go to the Projects page.
2. In the project list, select the project you want to delete and click Delete.
3. In the dialog, type the project ID, and then click Shut down to delete the project.

• Read the tutorial Structured data prediction using Cloud AI Platform to learn the machine learning model you use in this example.
• Learn more about AI on Google Cloud.
• Try out other Google Cloud features for yourself. Have a look at our tutorials.