This issue is a roadmap and checklist for ongoing work re-factoring DMatrix (see RFC #4354).

The first steps are to use a common interface to external data, unifying the way DMatrix objects are constructed and simplifying the process of adding new external data sources.

• Create adapter interface to external data, implement constructors for SimpleDMatrix (Demo of external data adapters #5044)
• Extend data adapters to SparsePageDMatrix (Use adapters for SparsePageDMatrix #5092)
• Extend data adapters to cudf data loading (Use dynamic types for array interface columns instead of templates #5108)
• Add support for cupy (Support dmatrix construction from cupy array #5206)
• Implement DMatrix slice via adapter (Implement slice via adapters #5198)

After the above all DMatrix constructors will be happening via adapters, missing value handling and use of threads will be consistent.

Then I plan to start reducing the number of classes associated with DMatrix.

• Remove SimpleCSRSource into SimpleDMatrix (Remove SimpleCSRSource #5315)
• Remove SparsePageSource into SparsePageDMatrix

The final goal is to save memory by constructing histogram matrices for the hist and gpu_hist algorithms directly from external data using adapters. We will need some discussion on the interface e.g. if a user wants to build the histogram DMatrix directly, specify an enum to the constructor indicating DMatrix type

• Develop interface for instantiating histogram DMatrix directly
• Build constructor for EllPack matrix directly using adapters

Lastly:

• Enable weighted sketching for DeviceQuantileDMatrix.