IPython magic for simple, organized, compressed and encrypted storage & transfer of files between notebooks.

The %vault magic provides a reproducible caching mechanism for variables exchange between notebooks. The cache is compressed, persistent and safe.

Differently to the builtin %store magic, the variables are stored in plain sight, in a zipped archive, so that they can be easily accessed for manual inspection, or for the use by other tools.

Let's open the vault (it will be created if not here yet):

%open_vault -p data/storage.zip

Generate some dummy dataset:

from pandas import DataFrame
from random import choice, randint
cities = ['London', 'Delhi', 'Tokyo', 'Lagos', 'Warsaw', 'Chongqing']
salaries = DataFrame([
    {'salary': randint(0, 100), 'city': choice(cities)}
    for i in range(10000)
])

And store it in the vault:

%vault store salaries in datasets

Stored salaries (None → 40CA7812) at Sunday, 08. Dec 2019 11:58

A short description is printed out (including a CRC32 hashsum and a timestamp) by default, but can be disabled by passing --timestamp False to %open_vault magic. Even more information enhancing the reproducibility is stored in the cell metadata.

We can now load the stored DataFrame in another (or the same) notebook:

%vault import salaries from datasets

Imported salaries (40CA7812) at Sunday, 08. Dec 2019 12:02

Thanks to (optional) memory optimizations we saved some RAM (87% as compared to unoptimized pd.read_csv() result). To track how many MB were saved use --report_memory_gain setting which will display memory optimization results below imports, for example:

Reduced memory usage by 87.28%, from 0.79 MB to 0.10 MB.

If we already have the salaries variable, we can use as, just like in the Python import system.

%vault import salaries from datasets as salaries_dataset

from pandas import read_csv
to_csv = lambda df: df.to_csv()
%vault store salaries in datasets with to_csv as salaries_csv
%vault import salaries_csv from datasets with read_csv

from pandas import read_excel
%vault import 'cars.xlsx' as cars_dataset with read_excel

More examples are available in the Examples.ipynb notebook, which can be run interactively in the browser.

Syntax:

• easy to understand in plain language (avoid abbreviations when possible),
• while intuitive for Python developers,
• ...but sufficiently different so that it would not be mistaken with Python constructs
  • for example, we could have %from x import y, but this looks very like normal Python; having %vault from x import y makes it sufficiently easy to distinguish
• star imports are better avoided, thus not supported
• as imports may be confusing if there is more than one

Reproducibility:

• promote good reproducible and traceable organization of files:
  • promote storage in plain text files and the use of DataFrame

    pickling is often an easy solution, but it can cause hurtful problems in prototyping phase (which is what notebooks are often used for): if you pickle you objects, then change the class definition and attempt to load your data again you are likely to fail severly; this is why the plain text files are the default option in this package (but pickling is supported too!).

  • print out a short hashsum and human-readable datetime (always in UTC),
  • while providing even more details in cell metadata
• allow to trace instances of the code being modified post execution

Security:

• think of it as a tool to minimize the damage in case of accidental git add of data files (even if those should have been elsewhere and .gitignored in the first place),
• or, as an additional layer of security for already anonymized data,
• but this tool is not aimed at facilitating the storage of highly sensitive data
• you have to set a password, or explicitly set --secure False to get rid of a security warning

Each operation will print out the timestamp and the CRC32 short checksum of the files involved. The timestamp of the operation is reported in the UTC timezone in a human-readable format.

This can be disabled by setting -t False or --timestamp False, however for the sake of reproducibility it is encouraged to keep this information visible in the notebook.

More precise information including the SHA256 cheksum (with a lower probability of collisions), and a full timestamp (to detect potential race condition errors in file write operations) are embedded in the metadata of the cell. You can disable this by setting --metadata False.

The exact command line is also stored in the metadata, so that if you accidentally modify the code cell without re-running the code, the change can be tracked down.

In order to enforce interoperability plain text files are used for pandas DataFrame and Series objects. Other variables are stores as pickle objects. The location of the storage archive on the disk defaults to storage.zip in the current directory, and can changed using %open_vault magic:

%open_vault -p custom_storage.zip

The encryption is not intended as a high security mechanism, but only as an additional layer of protection for already anonymized data.

The password to encrypt the storage archive is retrieved from the environmental variable, using a name provided in encryption_variable during the setup.

%open_vault -e ENV_STORAGE_KEY

Pandas DataFrames are by-default memory optimized by conversion of string variables to (ordered) categorical columns (pandas equivalent of R's factors/levels). Each string column will be tested for the memory improvement and the optimization will be only applied if it does reduce the memory usage.

The storage archive is conceptually similar to Hierarchical Data Format (e.g. HDF5) object - it contains:

• a hierarchy of files, and
• a metadata files

I believe that HDF may be the future, but this future is not here yet - numerous issues with the packages handling the HDF files, as well as low performance and compression rate prompted me to stay with a simple zip format now.

ZIP is a popular file format with known features and limitations - files can be password encrypted, while the file list is always accessible. This is okay given that the code of the project is assumed to be public, and only the files in the storage area are assumed to be of encrypted, increasing the security in case of unauthorized access.

As the limitations of the ZIP encryption are assumed to be a common knowledge, I hope that managing expectations of the level of security offered by this package will be easier.

Pre-requirements:

• Python 3.6+
• 7zip (16.02+) (see below for Ubuntu and Mac commands)

pip3 install data_vault

You can use p7zip packages from the default repositories:

sudo apt-get install -y p7zip-full

brew install p7zip

Installers for Windows can be downloaded from the 7-zip website.

Windows is not supported as it has known issues.