Switching from jupyter_server

Ganymede replaces the server, not the kernels. If you run jupyter lab today, your Python notebooks keep running on ipykernel, with pandas, matplotlib and the rest of your environment, once ipykernel's kernelspec is installed. What changes is the process between the frontend and the kernel: one binary instead of a Python environment with jupyter_server, jupyter_client, nbformat and traitlets in it.

This page lists what carries over, what is different, and what is missing, so you can judge the switch before making it.

What carries over

Notebooks	The same .ipynb files. Saves match Python's nbformat byte for byte, so they don't churn in git.
Kernels	Every installed kernelspec in the Jupyter data dirs and $JUPYTER_PATH, ipykernel's included.
Trust	The same notebook_secret and nbsignatures.db: a notebook trusted in Jupyter is trusted here.
Logins	The same jupyter_cookie_secret, so a browser stays logged in.
JupyterLab settings	~/.jupyter/lab/user-settings/ and workspaces/.
Clients	JupyterLab, VS Code's Jupyter extension, and other clients of the REST and websocket API. See JupyterLab and VS Code.
Headless runs	ganymede execute writes the same outputs as jupyter execute; ganymede trust replaces jupyter trust.

What is different

• Installing JupyterLab. ganymede lab install downloads JupyterLab's browser files from PyPI without running Python, instead of pip install jupyterlab. Prebuilt extensions install with ganymede lab extension install <package>.

• Kernels in an environment. jupyter_client also finds kernels under the Python environment it runs in (sys.prefix). Ganymede has no Python environment, so a kernel installed into a virtualenv or conda environment needs that prefix added: --kernelspec-root "$VIRTUAL_ENV/share/jupyter". A kernel installed with --user, as python -m ipykernel install --user does, is found as it is.

• The default kernel is --default-kernel, or else the first kernelspec found, not python3.

• Configuration is one TOML file instead of jupyter_server_config.py (CLI and configuration). The settings that carry over:

  jupyter_server	ganymede.toml
  ServerApp.ip, ServerApp.port	ip, port
  IdentityProvider.token	token
  ServerApp.root_dir	root_dir
  ServerApp.base_url	base_url
  ServerApp.allow_origin	allow_origins (a list)
  ServerApp.allow_credentials	allow_credentials
  ServerApp.allow_remote_access, ServerApp.local_hostnames	allow_remote_access, local_hostnames
  ServerApp.trust_xheaders	trust_xheaders
  ServerApp.cookie_secret_file	cookie_secret_file
  ServerApp.max_body_size	max_body_size
  MappingKernelManager.default_kernel_name	default_kernel
  MappingKernelManager.cull_idle_timeout	cull_idle_secs
  KernelRestarter.restart_limit	kernel_restart_limit
  FileCheckpoints.checkpoint_dir	checkpoint_dir
  NotebookNotary.data_dir	notary_dir
  GatewayClient.url, GatewayClient.auth_token	gateway_url, gateway_token

  Other settings have no equivalent. An unknown key in the file is an error, so a setting that doesn't carry over can't be ignored silently.

What is missing

Server extensions (the Python half of some JupyterLab extensions), nbconvert's template exports (HTML, PDF, LaTeX, slides; script and Markdown work), and JupyterHub's single-user mode. Terminals and language servers (JupyterLab and VS Code) and real-time collaboration (Live editing and document rooms) do work, with no Python behind them. Known gaps has the full list and the details. If you depend on one of these, stay on jupyter_server for now, or run it for that one job: both servers share kernels, notebooks and signatures, so they can run side by side.

TypeScript notebooks

Ganymede was built for TypeScript notebooks with cdp-kernel, with no Python installed at all. The JavaScript ecosystem for data work is smaller than Python's, and there is no single library with pandas' reach. If your work is pandas and matplotlib, keep it on ipykernel; Ganymede runs both kernels side by side, and you can move notebook by notebook.