FAQ

How to install new packages?

Tip

👉 Would you rather watch a short video tutorial? Check it our here: installing additional packages.

To install new packages, you should make use of environments. Simply build a new environment that contains your package and select it inside the pipeline editor. Installing packages is done using well known commands such as pip install and sudo apt-get install.

Note

💡 When updating an existing environment, the new environment will automatically be used inside the visual editor (and for your interactive pipeline runs). However, the JupyterLab kernel needs to be restarted if it was already running.

What not to do

Do not install new packages by running bash commands inside the Notebooks. This will require the packages to be installed every time you do a pipeline run, since the state of the kernel environment is ephemeral.

How to share code between steps?

Note

💡 This approach also works to share code between pipelines.

There are multiple answers to this question. One being that you can make that code into a package which you can then install in your environment, just like other packages such as numpy. Of course the development cycle would be highly reduced with this approach and so an alternative would be to add the files to the project directory directly and import them in your scripts.

For example, you could create a utils.py file in your project directory and use its functions from within your scripts by:

import utils

utils.transform(...)

How to minimize Orchest’s disk size?

To keep Orchest’s disk footprint to a minimal you can use the following best practices:

  • Are you persisting data to disk? Then write it to the /data directory instead of the project directory. Jobs create a snapshot (for reproducibility reasons) of your project directory and would copy data in your project directory for every pipeline run, consuming large amounts of storage. The smaller the size of your project directory, the smaller the size of your jobs.

  • Do you have many pipeline runs as part of jobs? You can configure your job to only retain a number of pipeline runs and automatically delete the older ones. Steps: (1) edit an existing job or create a new one, (2) go to pipeline runs, and (3) select auto clean-up.

How to use a GPU in Orchest?

Currently GPU support is not yet available. Coming soon!

How to skip notebook cells?

Notebooks facilitate an experimental workflow, meaning that there will be cells that should not be run when executing the notebook (from top to bottom). Since pipeline runs require your notebooks to be executable, Orchest provides an (pre-installed JupyterLab) extension to skip those cells.

To skip a cell during pipeline runs:

  1. Open JupyterLab.

  2. Go to the Property Inspector, this is the icon with the two gears all the way at the right.

  3. Select the cell you want to skip and give it a tag of: skip.

The cells with the skip tag are still runnable through JupyterLab, but when executing these notebooks as part of pipelines in Orchest they will not be run.

How to migrate to Kubernetes?

The moment we have moved to a Kubernetes backed Orchest version (and deprecated the Docker based version), we will update this section of the documentation to include steps on how to migrate your current deployment to a Kubernetes based one.

Just know that we are super excited to make the Kubernetes version available part of the open core and we are invested to provide a smooth migration experience 🔥

I’m getting ModuleNotFoundError: No module named exceptions even after declaring my dependencies

If you are getting weird ModuleNotFoundError exceptions for libraries that supposedly you declared already, there is a chance that you might have reinstalled ipykernel with pip, which causes a known incompatibility. There are two ways to fix this issue:

  1. Add python -m ipykernel install --sys-prefix at the end of your setup script, which restores the paths ipykernel needs to work.

  2. Use mamba (or conda) instead of pip to install your dependencies, which avoid this incompatibility.