Files and storage

User

Read this first

A container’s own filesystem is deleted when the job ends - whether it finished, expired, failed or was cancelled. Only what you wrote to a mounted bucket survives. There is no recovery.

The Files page

Storage cost, Your storage and Cost per hour at the top: the rate per terabyte-hour, what you are using, and what that is costing. Storage is billed continuously, not per job.

Below them, your storage buckets, each with an access badge:

Badge	Means
RW	Read and write
RO	Read only - mountable, but your job cannot write to it

Select a bucket to browse it, upload into it, and download from it.

Mounting a bucket into a job

1. In the submission form, Folders to mount → + ADD.
2. Pick the collection and the bucket.
3. MOUNT.
4. Adjust the container path if you want it somewhere other than the default.

Your job then reads and writes that path like any other directory, and what is there afterwards is still there.

Why the job form shows fewer buckets

This is the one genuinely confusing thing about storage, so it is worth stating plainly:

• The Files page lists every bucket you have access to.
• The job form lists every bucket the selected queue can reach.

A bucket lives on a storage, a storage is attached to particular queues, and a job can only mount storage its node can actually see. So a perfectly visible bucket can be unmountable on one queue and fine on another.

When the form says “No storage is attached to the … queue, so there is nothing to mount here”, your options are to pick a different queue, or to ask an administrator to attach that storage to the queue. See Storage administration.

Working with data

Getting data in. Upload through the Files page for small things. For large datasets, mount the bucket into a job and pull from inside the container - the node’s connection is usually much better than yours.

Getting results out. Write them to the mounted path, then download from the Files page after the job ends.

A practical layout. Keep datasets in a read-only bucket shared across the team, and write outputs to your own read-write one. Then a mistake in a job cannot destroy the dataset.

# inside the container
python train.py \
  --data   /datasets/imagenet \
  --output /homecatalog/runs/$(date +%F-%H%M)

Costs

Storage is charged per terabyte-hour, continuously, whether or not a job is running. A bucket you filled and forgot keeps costing. The Files page shows the current rate and your running total; the Tokens page shows what has actually been charged.