Skip to main content

Workspaces and persistent storage

A workspace is an isolated compute environment. It owns three things that matter to you daily: persistent storage, the image's preinstalled libraries, and a resource allocation.

What isolation means​

Two workspaces in the same organization do not share storage or installed packages. A library installed in one is absent from the other; a file written in one is invisible to the other.

That isolation is what lets one organization run separate projects, or separate production and experimental environments, without either interfering with the other.

Persistent storage​

Files in the workspace tree survive between sessions, except the scratch_work folder, which is temporary. This is where your work belongs.

# Persists: the workspace tree
df.to_parquet("/workspace/results/run-2026-08-30.parquet")

# Does NOT persist: session-local
df.to_parquet("/tmp/results.parquet")

The failure mode is an afternoon of output written to /tmp, gone after a restart. It happens to everyone once.

Local disk is not shared with the cluster

Workspace storage is visible to your session. Data that many workers need should be in object storage, which every compute node can reach. A path that works in your notebook is not automatically a path a worker can read.

Where data should live​

DataWhereWhy
Notebooks, scriptsWorkspace treePersists, and only your session needs it
Small reference filesWorkspace treeConvenient, and cheap to copy into tasks
Large inputs read by workersObject storageEvery node can reach it in parallel
Run outputsObject storageDurable, and readable outside the workspace

Libraries​

The scientific Python stack comes preinstalled: NumPy, pandas, SciPy, scikit-learn, PyTorch, plus common geospatial and imaging libraries.

For anything else:

%pip install some-package

This installs into your session's Python, next to the preinstalled libraries (Eugo's system site-packages under /usr/local). It lasts for that session: a new session starts from the image again, so keep the install line with your notebook and re-run it.

It also reaches your session only. The cluster's workers are a separate environment, and a package a task imports has to be sent to them; see getting extra packages onto the workers.

Not supported

TensorFlow, spaCy, and JAX are not supported on Eugo. Plan around PyTorch.

Resource allocation​

A workspace's limits come from its pricing plan. The plan governs concurrency as well as capacity: how much can run at once, not only how much total.

That distinction bites as a team grows. Ten people on a plan sized for two will queue, and the symptom looks like slowness rather than a limit.

Choosing the boundary​

Workspaces get split per team, per project, or per environment. Decide which before you create things, because storage and limits follow the boundary, and moving later means moving data. The workspace setup checklist walks through the decision in order.