Accessing Object Storage in GCP


This tutorial demonstrates how to access your Google Cloud Storage from the Neuro platform. You will create a new Neuro project, a new project in GCP, a service account, and a bucket that's accessible from a job on the Neuro platform.

Make sure you have CLI and cookiecutter installed.

Creating Neuro and GCP Projects

To create a new Neuro project and build an image, run:

$ cookiecutter gh:neuro-inc/cookiecutter-neuro-project --checkout release
$ cd <project-id>
$ neuro-flow build myimage

It's a good practice to limit the scope of access to a specific GCP project. To create a new GCP Project, run:

PROJECT_ID=${PWD##*/}  # name of the current directory
gcloud projects create $PROJECT_ID
gcloud config set project $PROJECT_ID

Make sure to set a billing account for your GCP project. See Creating and Managing Projects for details.

Creating a Service Account and Uploading an Account Key

First, create a service account for the job:

gcloud iam service-accounts create $SA_NAME  \
    --description "Neuro Platform Job Service Account" \
    --display-name "Neuro Platform Job"

See Creating and managing service accounts for more details.

Then, download the account key:

gcloud iam service-accounts keys create ~/$SA_NAME-key.json \
  --iam-account $SA_NAME@$

Make sure that the newly created key is located at ~/.

Create a new secret for the file:

$ neuro secret add gcp-key @~/$SA_NAME-key.json

Open .neuro/live.yaml, find the remote_debug section within jobs, and add the following lines at the end of remote_debug:

     secret_files: '["secret:gcp-key:/var/secrets/gcp.json"]'
     additional_env_vars: '{"GOOGLE_APPLICATION_CREDENTIALS": "/var/secrets/gcp.json"}'

Creating a Bucket and Granting Access

Now, create a new bucket. Remember: bucket names are globally unique (see more information on bucket naming conventions).

gsutil mb gs://$BUCKET_NAME/

Grant access to the bucket:

# Permissions for gsutil:
gsutil iam ch serviceAccount:$SA_NAME@$$PERM gs://$BUCKET_NAME

# Permissions for client APIs:
gsutil iam ch serviceAccount:$SA_NAME@$$PERM gs://$BUCKET_NAME


Create a file and upload it into Google Cloud Storage Bucket:

echo "Hello World" | gsutil cp - gs://$BUCKET_NAME/hello.txt

Change the default preset to cpu-small in .neuro/live.yamlto avoid consuming GPU for this test:

  preset: cpu-small

Run a development job and connect to the job's shell:

$ neuro-flow run remote_debug

In your job's shell, activate the service account for CLI:

gcloud auth activate-service-account --key-file $GOOGLE_APPLICATION_CREDENTIALS

And try to use gsutil to access your bucket:

gsutil cat gs://my-neuro-bucket-42/hello.txt

Please note that in remote_debug, train, and jupyter jobs the environment variable GOOGLE_APPLICATION_CREDENTIALS points to your key file. So you can use it to authenticate other libraries.

For instance, you can access your bucket via Python API provided by package google-cloud-storage:

>>> from import storage
>>> bucket = storage.Client().get_bucket("my-neuro-bucket-42")
>>> text = bucket.get_blob("hello.txt").download_as_string()
>>> print(text)
b'Hello World\n'

To close remote terminal session, press ^D or type exit.

Please don't forget to terminate your job when you don't need it anymore:

$ neuro-flow kill remote_debug

Last updated