datatable

This is a Python package for manipulating 2-dimensional tabular data structures (aka data frames). It is close in spirit to pandas or SFrame; however we put specific emphasis on speed and big data support. As the name suggests, the package is closely related to R's data.table and attempts to mimic its core algorithms and API.

Currently datatable is in the Beta stage and undergoing active development. Some of the features may still be missing. Python 3.6+ is required.

Project goals

datatable started in 2017 as a toolkit for performing big data (up to 100GB) operations on a single-node machine, at the maximum speed possible. Such requirements are dictated by modern machine-learning applications, which need to process large volumes of data and generate many features in order to achieve the best model accuracy. The first user of datatable was Driverless.ai.

The set of features that we want to implement with datatable is at least the following:

• Column-oriented data storage.

• Native-C implementation for all datatypes, including strings. Packages such as pandas and numpy already do that for numeric columns, but not for strings.

• Support for date-time and categorical types. Object type is also supported, but promotion into object discouraged.

• All types should support null values, with as little overhead as possible.

• Data should be stored on disk in the same format as in memory. This will allow us to memory-map data on disk and work on out-of-memory datasets transparently.

• Work with memory-mapped datasets to avoid loading into memory more data than necessary for each particular operation.

• Fast data reading from CSV and other formats.

• Multi-threaded data processing: time-consuming operations should attempt to utilize all cores for maximum efficiency.

• Efficient algorithms for sorting/grouping/joining.

• Expressive query syntax (similar to data.table).

• LLVM-based lazy computation for complex queries (code generated, compiled and executed on-the-fly).

• LLVM-based user-defined functions.

• Minimal amount of data copying, copy-on-write semantics for shared data.

• Use "rowindex" views in filtering/sorting/grouping/joining operators to avoid unnecessary data copying.

• Interoperability with pandas / numpy / pure python: the users should have the ability to convert to another data-processing framework with ease.

• Restrictions: Python 3.6+, 64-bit systems only.

Installation

On macOS, Linux and Windows systems installing datatable is as easy as

pip install datatable

On all other platforms a source distribution will be needed. For more information see Build instructions.

See also

• Build instructions
• Documentation