Skip to main content
Object Storage stores data for your ML/AI workloads. For example, you train a text-to-text model with a several petabytes dataset size. Here is how you can use Object Storage for it:
  1. Data storage: Upload your training data to a bucket. Each object in the bucket is a file with a piece of text data for training. Set this bucket as a data source for your training scripts.
  2. Checkpoint storage: During the model training and running experiments, set the scripts to upload each model checkpoint to the dedicated bucket.
  3. Model registry: Use Object Storage to distribute models across your services or consumers, or to leverage autoscaling during inference.
  4. Inference results storage: During the model inference, store the inference results in an Object Storage bucket.
Your data in Object Storage is stored as objects: files in various formats together with their metadata. To organize the objects, you create containers called buckets in your project and upload objects into them.

​
Buckets

Bucket is a container in Object Storage for storing files.

​
Naming

A bucket name should be unique across the region. If you try to create a bucket with a name that is already taken by another bucket in the region, you will get a message about it. The name must be 3 to 63 characters long and may contain lowercase Latin letters, numbers, dots and hyphens. It is used as a part of a URL for data access.

​
URLs

Use the following template to access a bucket: 
https://storage.<region_id>.nebius.cloud/<bucket_name>
In the URL above:
  • region_id: the ID of the region where the bucket is located.
  • bucket_name: the name of the bucket you want to access.
For example, the URL for the training-artifacts bucket located in eu-north1 (Finland) will be: 
https://storage.eu-north1.nebius.cloud/training-artifacts

​
Objects

Object is a file in Object Storage together with its metadata.

​
Naming

To avoid issues with your objects, follow these naming requirements:
  • Object keys can be up to 1024 bytes long, case sensitive.
  • Use UTF-8 alphanumeric characters, slashes (/) and certain special characters.

​
Keys, prefixes and object hierarchy

Each Object Storage object has an ID string called key. The key can contain prefixes that act similar to directories, organizing objects into groups. Buckets have flat structure: they have no directories in them, all objects in a bucket are at the same level of hierarchy. However, interfaces and tools supported by Object Storage, such as the AWS CLI, can emulate directories. If you upload a couple of objects with the same part of their IDs ending with /, they will be shown like they are in the same directory. The two uploaded files will be accessible as and .

​
Storage classes

Object Storage provides storage classes to control settings for stored objects. You can set the default storage class for each bucket and upload individual objects to specific classes. In addition, you can set up lifecycle rules to transition objects between classes.

​
URLs

Use the following templates to access an object: For example, the URL for the requests.txt object in the training-artifacts bucket located in eu-north1 (Finland) will be: 
https://storage.eu-north1.nebius.cloud/training-artifacts/requests.txt

​
Encryption

Nebius enforces strong encryption across all layers of data handling:
  • At rest: all objects in Object Storage are encrypted using AES-256 by default.
  • In transit: all communications use HTTPS (TLS 1.2 or higher).

​
Actions with buckets and objects

In your project, you can create buckets, upload objects to them and download objects from your buckets.