RAPIDS Docs

This site uses Just the Docs, a documentation theme for Jekyll.

Search

RSN 60 - Sunsetting cuxfilter after RAPIDS Release v26.06

Author RAPIDS TPM
Status

In Progress
Topic Platform Support Change
RAPIDS Version v26.06+
Created 22 May 2026
Updated N/A

Overview

RAPIDS v26.06 will be the final release to include updates for cuxfilter. After the v26.06 release, RAPIDS will stop publishing new cuxfilter packages, development of the cuxfilter repository will cease, and the repository will be frozen or archived with migration guidance.

cuxfilter helped users build GPU-accelerated, notebook-first, cross-filtered dashboards over large datasets by connecting cuDF-backed data to visualization libraries such as Panel, Bokeh, HoloViews, Datashader, and deck.gl.

The forward path for this workflow will be a skill-based replacement rather than a successor Python package. The skill will preserve the useful cuxfilter patterns and help users generate GPU-accelerated visual analytics directly with supported RAPIDS and Python visualization libraries.

Impact

cuxfilter v26.06 will remain the final maintained version. No new cuxfilter conda or pip packages will be published for RAPIDS releases after v26.06, and future RAPIDS releases will not provide compatibility updates, bug fixes, or API support for cuxfilter.

Beginning with the first RAPIDS release after v26.06, cuxfilter should be removed from RAPIDS package and release surfaces where applicable, including release manifests, install examples, documentation entry points, metapackage references, and container references.

Existing users may continue to use the v26.06 package by pinning compatible RAPIDS, CUDA, Python, and visualization-library versions. That path is intended only for existing workloads that cannot migrate immediately. It should not be used as the starting point for new dashboard or notebook development.

Migration guidance

There will not be a direct successor package that re-implements the cuxfilter API. New work should use the skill replacement and direct-library patterns:

  • Use cuDF for GPU dataframe loading, transformation, aggregation, and joins.
  • Use HoloViews, hvPlot, Datashader, and Panel for notebook-first visual exploration and linked selections.
  • Use Plotly Dash, Streamlit, Bokeh, or PyDeck when the desired output is a standalone application or a framework-specific dashboard.
  • Use pandas or Polars as a CPU fallback when a local GPU is unavailable.

The skill replacement is intended to capture the relevant workflow knowledge: GPU dataframe use, visual aggregation, linked selections, layout, and controls for fast exploration of large datasets.

Mapping common cuxfilter concepts

Previous cuxfilter concept Recommended replacement pattern
cuxfilter.DataFrame Load and transform data directly with cudf.DataFrame; convert only at explicit visualization boundaries when a library requires CPU data.
dashboard([...]) and preset layouts Compose the view with Panel, Dash, Streamlit, or another maintained dashboard framework.
Charts, widgets, and linked filters Use HoloViews/hvPlot/Datashader with link_selections, or framework-native callback/state patterns.
Graph and geospatial examples Use cuGraph or cuSpatial for GPU-side analytics, then visualize with Datashader, HoloViews, PyDeck, Bokeh, or Plotly.

Continued GPU visual analytics support

RAPIDS continues to support GPU-accelerated data preparation and analytics through projects such as cuDF, cuGraph, cuSpatial, and related libraries. The replacement skill will point users to those libraries and provide examples and templates for building accelerated visual analytics workflows without importing cuxfilter.