Managed JupyterLab / VS Code Server for data scientists — no infra knowledge required.

Profile model¶

Notebooks are templated:

Profile	Flavor	Use
notebook-cpu-small	std-2x8	Exploration, small data
notebook-cpu-large	mem-4x32	Pandas on real data
notebook-gpu-t4	gpu-t4-1x	Inference experiments
notebook-gpu-a100	gpu-a100-1x	Real training

Each profile defines image, flavor, max idle time, max session duration.

IAM¶

Role	Can do
notebook.user	Launch own notebooks from approved profiles
notebook.admin	Manage profiles, set quotas, view all sessions

notebook.user users see only their own sessions; admins see all (for cost & security audit).

Auto-stop¶

Default: 30-min idle stop, 8-h max-session. Tune per profile:

bash cd notebook profile set notebook-gpu-a100 \ --idle-stop 15min --max-session 6h

Auto-stop snapshots the workspace; user can resume.

Persistent workspaces¶

Each user gets a personal volume mounted at /home/user. Survives notebook stops; bills as block storage.

Pre-installed libraries¶

Base images include: - Python 3.11, R 4.3 - PyTorch 2.5, TensorFlow 2.16, JAX - pandas, numpy, scikit-learn, polars - jupyter, jupyterlab, vscode-server

Custom images supported for additional packages — push to the registry, set as the profile's image.

SSO and audit¶

Login via SSO/SAML/OIDC. Every session start, stop, and idle-stop is audit-logged.

For PII workloads, an admin can configure data-residency profiles that restrict the notebook's egress (no public internet; only internal endpoints).

Related¶

• Notebook — Operation
• Notebook — Troubleshooting

Metrics (admin view)¶

Metric	Notes
notebook.active_sessions	Current count, by profile
notebook.idle_sessions	Should auto-stop
notebook.gpu_hours.mtd	Track for cost
notebook.workspace_size_gb (per user)	Storage growth

Daily report by user emails to admins.

User onboarding¶

```bash cd notebook user invite --email jane@acme.com --quota gpu-a100=8h/day

Sends signup link¶

```

First-time users get a guided tour; an empty workspace; access to profiles their quota allows.

Custom image¶

Add team-specific packages:

Dockerfile FROM cloudigit/notebook-base:cuda-12.6 RUN pip install --no-cache-dir transformers==4.45 datasets accelerate

bash docker push registry.cloudigit.bd/acme-ml/notebook:v3 cd notebook profile create \ --name notebook-acme-llm \ --image registry.cloudigit.bd/acme-ml/notebook:v3 \ --flavor gpu-a100-1x

Quotas & cost¶

Per-user quotas across all profiles:

bash cd notebook quota set --user jane@acme.com --gpu-a100-hours-month 40

Soft warn at 80%; hard cut at 100%. Admin override available.

Sharing notebooks¶

• For collaboration: nbgrader / jupyterlab-collaboration extension (if enabled by admin)
• For results: export to HTML/PDF; share via internal portal
• For pipelines: extract code into a versioned repo

Don't use notebooks as the source-of-truth for production code.

Egress restrictions¶

For PII profiles:

bash cd notebook profile set notebook-pii \ --egress-allowlist 'internal-data.acme.local,registry.cloudigit.bd' \ --no-public-internet

Users can still use VPC-internal services; can't reach public package repos. Pre-bake required packages in the image.

Related¶

• Notebook — Administration
• Notebook — Troubleshooting

Notebook won't start¶

Cause	Check
Quota exhausted	cd notebook quota show --user <email>
No capacity for profile's flavor	Wait, or pick a smaller profile
Custom image broken (won't pull)	See Registry troubleshooting
Workspace volume offline	Ticket

cd notebook session diagnose --session <id> shows the failure reason.

Kernel keeps dying¶

Reason	Action
OOM (out of RAM)	Bigger profile, or work on a subset
GPU OOM	Bigger GPU profile, smaller batch
Long-running cell killed by idle-stop	Tune idle-stop on the profile
Network drop during pip install	Retry; check egress config

"No space left on device"¶

Workspace volume full:

bash df -h /home/user du -sh /home/user/*

Common culprits: - ~/.cache/huggingface (model downloads accumulate) - Conda envs cluttering - Untracked experiment outputs

Workspace volumes can be resized by admins.

VS Code extensions missing¶

VS Code Server runs the extensions in the server, not your local IDE. The pre-installed image has a baseline; missing extensions must be added by an admin to the custom image (not per-user via the marketplace, which is blocked in egress-restricted profiles).

Slow notebook UI¶

Symptom	Cause
Slow to render long outputs	Notebook with > 10MB of output; truncate / clear outputs
Slow file browser	Workspace has 100k+ small files; clean up
Slow Python imports	Cold cache; pre-import in profile startup script

Auto-stop killed my work¶

Workspace is persisted; code is saved (Jupyter auto-saves every minute). Notebook kernel state (variables in memory) is lost.

Mitigations: - Use %store to persist variables across sessions - Save intermediate results to disk frequently - Use a longer idle-stop for genuinely long-running computations

SSO login loop¶

Failed to authenticate: invalid_state

Cookie / session issue. Usually: - Browser blocking third-party cookies — allow for the IdP domain - Clock skew on user's device > 5 min — sync NTP

Related¶

• Notebook — Administration
• Notebook — Operation