2026)¶

🎉 Highlights¶

This release, we are excited to introduce Bodo’s CPU–GPU hybrid execution model available via conda (currently only supports Linux/Python 3.14):

conda install -c bodo.ai -c rapidsai -c nvidia "bodo=*=*cuda" --no-channel-priority

With the CUDA-enabled variant of the Bodo package, supported DataFrames operations will automatically run on all available GPUs, while unsupported operations seamlessly fall back to CPU.

The CUDA-enabled Bodo package uses CUDA-aware OpenMPI for GPU communication. Before running CPU–GPU hybrid code, you must first set OpenMPI’s communication layer to UCX:

export OMPI_MCA_pml=ucx

✨ CPU–GPU hybrid execution¶

The following types of operations will run on GPUs: - Parquet read (local filesystem, S3, Google Cloud Storage, HDFS, Azure Data Lake)

Parquet write (local filesystem, S3, Google Cloud Storage, HDFS, Azure Data Lake)
Most kinds of joins
Vectorized filters and projections (except UDFs)
Column arithmetic and boolean ops (except UDFs)
GroupBy aggregations: sum, count, mean, min, max, var, std, size, skew, nunique
Series reductions: sum, product, count, mean, min, max
drop_duplicates, concat, Series.isin

🏎️ Performance Improvements¶

Improved join ordering.

🛠️ Infrastructure¶

Added Docker ARM build.

⚙️ Dependency Changes¶

Upgraded DuckDB planner.