kartothek - manage tabular data in object stores




Jun 10, 2021

Kartothek is a Python library to manage (create, read, update, delete) large amounts of tabular data in a blob store. It stores data as datasets, which it presents as pandas DataFrames to the user. Datasets are a collection of files with the same schema that reside in a blob store. Kartothek uses a metadata definition to handle these datasets efficiently. For distributed access and manipulation of datasets, Kartothek offers a Dask interface (kartothek.io.dask).

Storing data distributed over multiple files in a blob store (S3, ABS, GCS, etc.) allows for a fast, cost-efficient and highly scalable data infrastructure. A downside of storing data solely in an object store is that the storages themselves give little to no guarantees beyond the consistency of a single file. In particular, they cannot guarantee the consistency of your dataset. If we demand a consistent state of our dataset at all times, we need to track the state of the dataset. Kartothek frees us from having to do this manually.

The kartothek.io module provides building blocks to create and modify these datasets in data pipelines. Kartothek handles I/O, tracks dataset partitions and selects subsets of data transparently.

To get started, have a look at our Getting Started guide, head to the description of the Specification or head straight to the API documentation API.

What is a (real) Kartothek?

A Kartothek (or more modern: Zettelkasten/Katalogkasten) is a tool to organize (high-level) information extracted from a source of information.