Terraform创建Cloud Dataflow Workbench实例后Notebook无法启动，需特定配置？

通过Terraform部署Google Cloud Dataflow Workbench实例时，使用容器镜像块搭配Notebook实例资源搭建Apache Beam环境（用户托管Notebook环境），尝试从控制台打开Notebook出现524错误。查看后端VM错误日志发现，notebook-collection agent无法连接Jupyter API服务器，服务器未启动。

注：手动创建相同实例可正常启动Notebook，排除防火墙、网络、服务账号/角色等问题。

疑问：是否需要在Terraform的容器镜像块中配置特定参数，才能让Notebook正常启动？

Terraform代码片段：

resource "google_notebooks_instance" "instance" {
  name = "notebooks-instance"
  location = "us-central1-a"
  machine_type = "e2-medium"

   container_images {
        repository = "gcr.io/deeplearning-platform-release/beam-notebooks"
        tag = "latest"
      }

  service_account = "my@service-account.com"

  install_gpu_driver = true
  boot_disk_type = "PD_SSD"
  boot_disk_size_gb = 110

  no_public_ip = true
  no_proxy_access = false

  network = data.google_compute_network.my_network.id
  subnet = data.google_compute_subnetwork.my_subnetwork.id

  labels = {
    k = "val"
  }

  metadata = {
    proxy-mode=service-account
    terraform = "true"
  }
}

需要在container_images块中添加container_image_type参数并设置为USER_MANAGED，具体修改如下：

resource "google_notebooks_instance" "instance" {
  name = "notebooks-instance"
  location = "us-central1-a"
  machine_type = "e2-medium"

  container_images {
    repository        = "gcr.io/deeplearning-platform-release/beam-notebooks"
    tag               = "latest"
    container_image_type = "USER_MANAGED" # 新增关键参数
  }

  service_account = "my@service-account.com"

  install_gpu_driver = true
  boot_disk_type = "PD_SSD"
  boot_disk_size_gb = 110

  no_public_ip = true
  no_proxy_access = false

  network = data.google_compute_network.my_network.id
  subnet = data.google_compute_subnetwork.my_subnetwork.id

  labels = {
    k = "val"
  }

  metadata = {
    proxy-mode=service-account
    terraform = "true"
  }
}

原因说明 #

使用自定义容器镜像搭建用户托管Notebook环境时，Terraform的google_notebooks_instance资源必须明确指定container_image_type为USER_MANAGED。默认情况下如果缺失该参数，资源会采用GCP托管镜像的启动流程，和用户托管容器的初始化逻辑不匹配，导致Jupyter服务无法正常启动。手动创建实例时，控制台会自动为用户托管类型配置该参数，所以不会出现问题。

内容的提问来源于stack exchange，提问作者Muralidharan