Contents

Getting Started

1. Create a Microsoft account if you don’t already have one.

   

2. Go to the Microsoft Azure Notebooks web site and sign in.

   

3. Select “Libraries” in the navigation bar (libraries are groups of related notebooks) and create a new library named “Sandbox”.

   

4. Create a new notebook named “My First Notebook.ipynb” in the Sandbox library, or upload an existing one. For this article, I will use a new Python 2.7 notebook.

   

5. Start the notebook.

   

The Azure Platform

To explore the Azure plaform hosting the Jupyter notebook, we will issue some shell commands; the simplest way to do that if from within a Python notebook is to type the command in a cell, prefixed with an exclamation point¹.

First of all, Azure notebooks are hosted on Linux (Debian-based) machine:

>>> import platform
>>> platform.system()
'Linux'
>>> platform.platform()
'Linux-4.4.0-81-generic-x86_64-with-debian-stretch-sid'

The distribution used is the latest LTS version of Ubuntu: Xenial Xerus.

>>> !cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=16.04
DISTRIB_CODENAME=xenial
DISTRIB_DESCRIPTION="Ubuntu 16.04.2 LTS"

The notebook is actually running within a Docker container:

>>> import os.path
>>> os.path.isfile("/.dockerenv")
True

What it means concretely is that when you start – or restart – a notebook, you are likely to wait for a couple of seconds while Azure is provisioning a new container. As far as I can tell, container instances are shared between notebooks in the same library, but not across libraries.

The processor specs are:

>> !lscpu | grep "Model name"
Model name: Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz

PassMark gives this CPU a mark of 16.904; this is far better than the laptop I am working on, so I guess that performance is uniquely to be an issue.

Data Management

If you use the Jupyter Notebook App, which executes the notebooks on your computer, or if you have deployed a JupyterHub server, the input data that you use for your notebooks, the ouput data that they may produce and the notebooks themselves are in the same filesystem, probably organized into directories, one for each projects.

Things are different for Azure notebooks, where notebooks and data are handled separately. Notebook files (with the .ipynb extension) are stored permanently and associated to your Microsoft account; they can be managed from the Libraries view and/or from the dashboard. You can also download/upload them if you loke but this is not mandatory. However, you will not find them in the filesystem accessible in the notebook ².

Your data files on the other hand – anything that is in the notebook filesystem – is ephemeral and will be lost when your library container is shut down. Consequently:

• If your notebook requires some input data, you need to check that it’s available before you execute the notebook, or “rehydrate” your filesystem, upload this data again.

• If your notebook produces some output data that you want to keep, you need to download it.

For both steps there are several options available.

• You may upload/download your data manually from/to your computer: use the “Data” menu in the notebook navigation panel. Data may also be uploaded from a Dropbox account.

• To compress the files in the current directory into a data.zip archive – but anaconda Python distributions and hidden files – and upload the archive to transfer.sh, type in a cell:

  >>> !zip -q -r data.zip . -x "anaconda*" ".*"
  >>> !curl --upload-file data.zip https://transfer.sh/data.zip

  The last command prints an address, something like:

  https://transfer.sh/yKwUl/data.zip

  This is where your archive is located (and will be for 14 days). To download and unzip this archive, type:

  >>> !curl https://transfer.sh/yKwUl/data.zip -o data.zip
  >>> !unzip -o data.zip

  Note that these commands work in Azure notebooks because https://transfer.sh is explicitly whitelisted (see Networking), so use specifically this service; other file sharing sites probably won’t work.

Software Packages

By default, the Azure notebook platform comes with a large set of pre-installed software packages provided by Anaconda, a Python distribution popular in numerical analysis and data science circles. Actually, three different versions of the Anaconda distributions are installed:

>>> !ls
anaconda2_410  anaconda3_410  anaconda3_431

Each distribution supports a different version of Python (at the time of writing: Python 2.7.11, 3.5.1 and 3.6.0).

To see for yourself the list of installed packaged, type:

>>> !conda list
# packages in environment at /home/nbcommon/anaconda2_410:
#
_nb_ext_conf              0.2.0                    py27_0  
adal                      0.4.6                     <pip>
alabaster                 0.7.8                    py27_0  
altair                    1.2.0                     <pip>
anaconda                  custom                   py27_0  
anaconda-client           1.4.0                    py27_0  
anaconda-navigator        1.2.1                    py27_0
... 

The full list is rather large; refer to the appendix if you are interested. The list is also compared with the default set of packages in the Anaconda distribution. There are generally more packages in the Azure notebook platform; some of them are obviously Azure-specific. Additionally, a package that would be missing from the Azure platform – for example wrapt – can easily be installed, either with

>>> !conda install -y wrapt

or – as long as it’s available on PyPI – with

>>> !pip install wrapt

Note that these installations are performed as a user, not at the system level: you are merely nbuser and you don’t have administrator rights in the Azure container. In particular, you won’t be able to apt-get install your way out of missing software.

Networking

Appendix – Conda Packages