Developer Guide

Note

At present, this guide only covers the main cuDF library. In the future, it may be expanded to also cover dask_cudf, cudf_kafka, and custreamz.

cuDF is a GPU-accelerated, Pandas-like DataFrame library. Under the hood, all of cuDF’s functionality relies on the CUDA-accelerated libcudf C++ library. Thus, cuDF’s internals are designed to efficiently and robustly map pandas APIs to libcudf functions. For more information about the libcudf library, a good starting point is the developer guide.

This document assumes familiarity with the overall contributing guide. The goal of this document is to provide more specific guidance for Python developers. It covers the structure of the Python code and discusses best practices. Additionally, it includes longer sections on more specific topics like testing and benchmarking.

• Library Design
  • The Frame layer
  • The Column layer
  • The Cython layer
  • Putting It All Together
  • Copy-on-write
• Contributing Guide
  • Directory structure and file naming
  • Code style
  • Deprecating and removing code
  • pandas compatibility
  • Python vs Cython
  • Exception handling
• Writing documentation
  • Docstrings
  • Published documentation
  • Comparing to pandas
  • Writing documentation pages
  • Building documentation
  • Documenting cuDF internals
• Testing cuDF
  • Tooling
  • Test organization
  • Test contents
• Benchmarking cuDF
  • Benchmark organization
  • Running benchmarks
  • Benchmark contents
  • Comparing to pandas
  • Testing benchmarks
  • Profiling
  • Advanced Topics
• Options
• pylibcudf
  • General Design Principles
  • Relationship to libcudf
  • Testing
  • Miscellaneous Notes
• cudf.pandas
  • fast-slow proxy mechanism
  • debugging cudf.pandas