numbagg

What it is and what it does

Numbagg is a library of fast aggregation and moving-window functions for N-dimensional NumPy arrays, written in Numba and exposed as generalized ufuncs. It includes standard reductions (nansum, nanmean, nanstd, nanvar), grouping operations (group_nansum, group_nanmean, etc.), moving-window statistics (move_mean, move_std, move_corr), and fill operations (ffill, bfill). The library is designed to outperform pandas and NumPy on multi-core systems through Numba's JIT compilation and parallelization, though single-threaded performance is comparable to NumPy and the first call to any function incurs JIT overhead.

It targets scientific and data-analysis workflows where performance on large arrays matters and flexible axis handling across many dimensions is needed. The main trade-off is that functions are JIT-compiled on first use, so initial calls are slower; subsequent calls are fast. Dependencies are minimal (numpy and numba), and the package supports Python 3.10 and later.

Use it for:

• Compute moving averages and standard deviations on time-series data with NaN handling faster than pandas.
• Apply grouped aggregations (sum, mean, count) on multi-dimensional scientific data.
• Calculate rolling correlations and covariances on financial or climate datasets.
• Forward-fill or backward-fill missing values in large arrays with better performance than pandas.
• Parallelize custom reductions across multiple cores on arrays too large for pure NumPy.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Numbagg provides fast N-dimensional aggregation and moving-window functions compiled with Numba, including reductions (nansum, nanmean, nanstd), grouping operations, and forward/backward fill, optimized for multi-core parallelization.

Yes. Numbagg is worth installing if you work with large N-dimensional NumPy arrays and need fast aggregations or moving-window functions, especially on multi-core systems. Low install friction, active maintenance, no known vulnerabilities, and permissive licensing make it a safe choice. The JIT compilation overhead on first call is a minor drawback but typical for Numba-based libraries and acceptable for most scientific workflows.

Install

numbagg on PyPI

pip install numbagg

uv add numbagg

poetry add numbagg

Installing numbagg

Before you install

Low install friction with a pure Python wheel and only two runtime dependencies (numpy and numba). Actively maintained with a recent release and no known vulnerabilities.

License in practice

BSD 3-Clause license with portions from Bottleneck (Simplified BSD). Permissive for commercial and private use; requires attribution and license notice in redistributions.

Quickstart

import numbagg
import numpy as np

arr = np.array([1.0, 2.0, np.nan, 4.0])
result = numbagg.nansum(arr)
# or for moving window:
window_result = numbagg.move_mean(arr, window=2)

Requires Python 3.10 or later. First execution of any function triggers JIT compilation, which adds startup latency.

Verify before relying

• Whether all listed functions (move_corr, group_nansum, etc.) are fully stable or still experimental.
• Exact memory overhead of JIT compilation relative to raw numpy for small arrays.
• Whether axis parameter behavior is documented for >3 dimensions in current release.

Package facts

Evidence: numbagg-0.9.4-py3-none-any.whl

Tags

More Scientific/Engineering packages