Numbagg: Fast N-dimensional aggregation functions with Numba

Fast, flexible N-dimensional array functions written with Numba and NumPy's generalized ufuncs.

Why use numbagg?

Performance

• Outperforms pandas
  • On a single core, 2-10x faster for moving window functions, 1-2x faster for aggregation and grouping functions
  • When parallelizing with multiple cores, 4-30x faster
• Outperforms bottleneck on multiple cores
  • On a single core, matches bottleneck
  • When parallelizing with multiple cores, 3-7x faster
• Outperforms numpy on multiple cores
  • On a single core, matches numpy
  • When parallelizing with multiple cores, 5-15x faster
• ...though numbagg's functions are JIT compiled, so the first run is much slower

Versatility

• More functions (though bottleneck has some functions we don't have, and pandas' functions have many more parameters)
• Functions work for >3 dimensions. All functions take an arbitrary axis or tuple of axes to calculate over
• Written in numba — way less code, simple to inspect, simple to improve

Functions & benchmarks

Summary benchmark

Two benchmarks summarize numbagg's performance — the first with a 1D array of 10M elements without parallelization, and a second with a 2D array of 100x10K elements with parallelization. Numbagg's relative performance is much higher where parallelization is possible. A wider range of arrays is listed in the full set of benchmarks below.

The values in the table are numbagg's performance as a multiple of other libraries for a given shaped array calculated over the final axis. (so 1.00x means numbagg is equal, higher means numbagg is faster.)

func	1D pandas	1D bottleneck	1D numpy	2D pandas	2D bottleneck	2D numpy
bfill	1.17x	1.18x	n/a	12.24x	4.36x	n/a
ffill	1.17x	1.12x	n/a	12.76x	4.34x	n/a
group_nanall	1.44x	n/a	n/a	10.84x	n/a	n/a
group_nanany	1.20x	n/a	n/a	5.25x	n/a	n/a
group_nanargmax	2.88x	n/a	n/a	9.89x	n/a	n/a
group_nanargmin	2.82x	n/a	n/a	9.96x	n/a	n/a
group_nancount	1.01x	n/a	n/a	4.70x	n/a	n/a
group_nanfirst	1.39x	n/a	n/a	11.80x	n/a	n/a
group_nanlast	1.16x	n/a	n/a	5.36x	n/a	n/a
group_nanmax	1.14x	n/a	n/a	5.22x	n/a	n/a
group_nanmean	1.19x	n/a	n/a	5.64x	n/a	n/a
group_nanmin	1.13x	n/a	n/a	5.26x	n/a	n/a
group_nanprod	1.15x	n/a	n/a	4.95x	n/a	n/a
group_nanstd	1.18x	n/a	n/a	5.03x	n/a	n/a
group_nansum_of_squares	1.35x	n/a	n/a	8.11x	n/a	n/a
group_nansum	1.21x	n/a	n/a	5.95x	n/a	n/a
group_nanvar	1.19x	n/a	n/a	5.65x	n/a	n/a
move_corr	19.04x	n/a	n/a	92.48x	n/a	n/a
move_cov	14.58x	n/a	n/a	71.61x	n/a	n/a
move_exp_nancorr	6.73x	n/a	n/a	35.30x	n/a	n/a
move_exp_nancount	2.35x	n/a	n/a	10.56x	n/a	n/a
move_exp_nancov	5.77x	n/a	n/a	31.75x	n/a	n/a
move_exp_nanmean	2.03x	n/a	n/a	11.07x	n/a	n/a
move_exp_nanstd	1.89x	n/a	n/a	10.07x	n/a	n/a
move_exp_nansum	1.88x	n/a	n/a	9.70x	n/a	n/a
move_exp_nanvar	1.82x	n/a	n/a	9.71x	n/a	n/a
move_mean	3.82x	0.87x	n/a	16.61x	4.01x	n/a
move_std	5.96x	1.29x	n/a	24.52x	6.04x	n/a
move_sum	3.80x	0.83x	n/a	15.95x	3.70x	n/a
move_var	5.78x	1.27x	n/a	25.41x	5.85x	n/a
nanargmax1	2.45x	1.00x	n/a	2.16x	1.00x	n/a
nanargmin1	2.19x	1.01x	n/a	2.05x	1.02x	n/a
nancount	1.40x	n/a	1.06x	11.00x	n/a	4.16x
nanmax1	3.26x	1.00x	0.11x	3.62x	3.24x	0.11x
nanmean	2.42x	0.98x	2.83x	13.58x	4.54x	13.13x
nanmin1	3.27x	1.00x	0.11x	3.62x	3.24x	0.11x
nanquantile	0.94x	n/a	0.78x	5.45x	n/a	5.01x
nanstd	1.50x	1.51x	2.75x	8.29x	7.35x	13.27x
nansum	2.28x	0.97x	2.52x	17.71x	6.24x	16.05x
nanvar	1.50x	1.49x	2.81x	8.18x	6.97x	13.32x

Full benchmarks

func	shape	size	pandas	bottleneck	numpy	numbagg	pandas_ratio	bottleneck_ratio	numpy_ratio	numbagg_ratio
bfill	(1000,)	1000	0ms	0ms	n/a	0ms	1.59x	0.03x	n/a	1.00x
	(10000000,)	10000000	20ms	20ms	n/a	17ms	1.17x	1.18x	n/a	1.00x
	(100, 100000)	10000000	57ms	20ms	n/a	5ms	12.24x	4.36x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	21ms	n/a	5ms	n/a	4.40x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	248ms	n/a	44ms	n/a	5.70x	n/a	1.00x
ffill	(1000,)	1000	0ms	0ms	n/a	0ms	1.53x	0.02x	n/a	1.00x
	(10000000,)	10000000	20ms	19ms	n/a	17ms	1.17x	1.12x	n/a	1.00x
	(100, 100000)	10000000	56ms	19ms	n/a	4ms	12.76x	4.34x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	19ms	n/a	4ms	n/a	4.33x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	219ms	n/a	42ms	n/a	5.25x	n/a	1.00x
group_nanall	(1000,)	1000	0ms	n/a	n/a	0ms	1.79x	n/a	n/a	1.00x
	(10000000,)	10000000	68ms	n/a	n/a	47ms	1.44x	n/a	n/a	1.00x
	(100, 100000)	10000000	17ms	n/a	n/a	2ms	10.84x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	1ms	n/a	n/a	n/a	1.00x
group_nanany	(1000,)	1000	0ms	n/a	n/a	0ms	1.78x	n/a	n/a	1.00x
	(10000000,)	10000000	68ms	n/a	n/a	56ms	1.20x	n/a	n/a	1.00x
	(100, 100000)	10000000	18ms	n/a	n/a	3ms	5.25x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanargmax	(1000,)	1000	1ms	n/a	n/a	0ms	17.60x	n/a	n/a	1.00x
	(10000000,)	10000000	171ms	n/a	n/a	59ms	2.88x	n/a	n/a	1.00x
	(100, 100000)	10000000	40ms	n/a	n/a	4ms	9.89x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	4ms	n/a	n/a	n/a	1.00x
group_nanargmin	(1000,)	1000	1ms	n/a	n/a	0ms	17.56x	n/a	n/a	1.00x
	(10000000,)	10000000	166ms	n/a	n/a	59ms	2.82x	n/a	n/a	1.00x
	(100, 100000)	10000000	41ms	n/a	n/a	4ms	9.96x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	4ms	n/a	n/a	n/a	1.00x
group_nancount	(1000,)	1000	0ms	n/a	n/a	0ms	1.68x	n/a	n/a	1.00x
	(10000000,)	10000000	56ms	n/a	n/a	55ms	1.01x	n/a	n/a	1.00x
	(100, 100000)	10000000	15ms	n/a	n/a	3ms	4.70x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanfirst	(1000,)	1000	0ms	n/a	n/a	0ms	1.88x	n/a	n/a	1.00x
	(10000000,)	10000000	63ms	n/a	n/a	45ms	1.39x	n/a	n/a	1.00x
	(100, 100000)	10000000	15ms	n/a	n/a	1ms	11.80x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	1ms	n/a	n/a	n/a	1.00x
group_nanlast	(1000,)	1000	0ms	n/a	n/a	0ms	1.87x	n/a	n/a	1.00x
	(10000000,)	10000000	62ms	n/a	n/a	53ms	1.16x	n/a	n/a	1.00x
	(100, 100000)	10000000	15ms	n/a	n/a	3ms	5.36x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	2ms	n/a	n/a	n/a	1.00x
group_nanmax	(1000,)	1000	0ms	n/a	n/a	0ms	1.89x	n/a	n/a	1.00x
	(10000000,)	10000000	66ms	n/a	n/a	57ms	1.14x	n/a	n/a	1.00x
	(100, 100000)	10000000	17ms	n/a	n/a	3ms	5.22x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanmean	(1000,)	1000	0ms	n/a	n/a	0ms	1.81x	n/a	n/a	1.00x
	(10000000,)	10000000	67ms	n/a	n/a	57ms	1.19x	n/a	n/a	1.00x
	(100, 100000)	10000000	19ms	n/a	n/a	3ms	5.64x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanmin	(1000,)	1000	0ms	n/a	n/a	0ms	1.84x	n/a	n/a	1.00x
	(10000000,)	10000000	66ms	n/a	n/a	58ms	1.13x	n/a	n/a	1.00x
	(100, 100000)	10000000	17ms	n/a	n/a	3ms	5.26x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanprod	(1000,)	1000	0ms	n/a	n/a	0ms	1.86x	n/a	n/a	1.00x
	(10000000,)	10000000	63ms	n/a	n/a	55ms	1.15x	n/a	n/a	1.00x
	(100, 100000)	10000000	16ms	n/a	n/a	3ms	4.95x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanstd	(1000,)	1000	0ms	n/a	n/a	0ms	1.73x	n/a	n/a	1.00x
	(10000000,)	10000000	70ms	n/a	n/a	59ms	1.18x	n/a	n/a	1.00x
	(100, 100000)	10000000	20ms	n/a	n/a	4ms	5.03x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	4ms	n/a	n/a	n/a	1.00x
group_nansum	(1000,)	1000	0ms	n/a	n/a	0ms	1.89x	n/a	n/a	1.00x
	(10000000,)	10000000	67ms	n/a	n/a	56ms	1.21x	n/a	n/a	1.00x
	(100, 100000)	10000000	19ms	n/a	n/a	3ms	5.95x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nanvar	(1000,)	1000	0ms	n/a	n/a	0ms	1.71x	n/a	n/a	1.00x
	(10000000,)	10000000	69ms	n/a	n/a	58ms	1.19x	n/a	n/a	1.00x
	(100, 100000)	10000000	20ms	n/a	n/a	4ms	5.65x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
group_nansum_of_squares	(1000,)	1000	0ms	n/a	n/a	0ms	2.36x	n/a	n/a	1.00x
	(10000000,)	10000000	75ms	n/a	n/a	55ms	1.35x	n/a	n/a	1.00x
	(100, 100000)	10000000	26ms	n/a	n/a	3ms	8.11x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	3ms	n/a	n/a	n/a	1.00x
move_corr	(1000,)	1000	0ms	n/a	n/a	0ms	10.85x	n/a	n/a	1.00x
	(10000000,)	10000000	909ms	n/a	n/a	48ms	19.04x	n/a	n/a	1.00x
	(100, 100000)	10000000	869ms	n/a	n/a	9ms	92.48x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	9ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	79ms	n/a	n/a	n/a	1.00x
move_cov	(1000,)	1000	0ms	n/a	n/a	0ms	10.05x	n/a	n/a	1.00x
	(10000000,)	10000000	623ms	n/a	n/a	43ms	14.58x	n/a	n/a	1.00x
	(100, 100000)	10000000	603ms	n/a	n/a	8ms	71.61x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	8ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	72ms	n/a	n/a	n/a	1.00x
move_mean	(1000,)	1000	0ms	0ms	n/a	0ms	1.84x	0.03x	n/a	1.00x
	(10000000,)	10000000	120ms	27ms	n/a	31ms	3.82x	0.87x	n/a	1.00x
	(100, 100000)	10000000	113ms	27ms	n/a	7ms	16.61x	4.01x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	27ms	n/a	7ms	n/a	3.96x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	296ms	n/a	58ms	n/a	5.08x	n/a	1.00x
move_std	(1000,)	1000	0ms	0ms	n/a	0ms	2.21x	0.08x	n/a	1.00x
	(10000000,)	10000000	178ms	39ms	n/a	30ms	5.96x	1.29x	n/a	1.00x
	(100, 100000)	10000000	157ms	39ms	n/a	6ms	24.52x	6.04x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	39ms	n/a	7ms	n/a	5.88x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	411ms	n/a	58ms	n/a	7.13x	n/a	1.00x
move_sum	(1000,)	1000	0ms	0ms	n/a	0ms	1.81x	0.02x	n/a	1.00x
	(10000000,)	10000000	121ms	26ms	n/a	32ms	3.80x	0.83x	n/a	1.00x
	(100, 100000)	10000000	113ms	26ms	n/a	7ms	15.95x	3.70x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	26ms	n/a	7ms	n/a	3.59x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	281ms	n/a	59ms	n/a	4.77x	n/a	1.00x
move_var	(1000,)	1000	0ms	0ms	n/a	0ms	2.04x	0.08x	n/a	1.00x
	(10000000,)	10000000	168ms	37ms	n/a	29ms	5.78x	1.27x	n/a	1.00x
	(100, 100000)	10000000	161ms	37ms	n/a	6ms	25.41x	5.85x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	37ms	n/a	6ms	n/a	5.85x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	398ms	n/a	56ms	n/a	7.07x	n/a	1.00x
move_exp_nancorr	(1000,)	1000	0ms	n/a	n/a	0ms	7.27x	n/a	n/a	1.00x
	(10000000,)	10000000	464ms	n/a	n/a	69ms	6.73x	n/a	n/a	1.00x
	(100, 100000)	10000000	471ms	n/a	n/a	13ms	35.30x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	13ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	111ms	n/a	n/a	n/a	1.00x
move_exp_nancount	(1000,)	1000	0ms	n/a	n/a	0ms	2.04x	n/a	n/a	1.00x
	(10000000,)	10000000	77ms	n/a	n/a	33ms	2.35x	n/a	n/a	1.00x
	(100, 100000)	10000000	69ms	n/a	n/a	7ms	10.56x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	6ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	59ms	n/a	n/a	n/a	1.00x
move_exp_nancov	(1000,)	1000	0ms	n/a	n/a	0ms	7.07x	n/a	n/a	1.00x
	(10000000,)	10000000	298ms	n/a	n/a	52ms	5.77x	n/a	n/a	1.00x
	(100, 100000)	10000000	333ms	n/a	n/a	10ms	31.75x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	10ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	87ms	n/a	n/a	n/a	1.00x
move_exp_nanmean	(1000,)	1000	0ms	n/a	n/a	0ms	1.40x	n/a	n/a	1.00x
	(10000000,)	10000000	67ms	n/a	n/a	33ms	2.03x	n/a	n/a	1.00x
	(100, 100000)	10000000	74ms	n/a	n/a	7ms	11.07x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	7ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	60ms	n/a	n/a	n/a	1.00x
move_exp_nanstd	(1000,)	1000	0ms	n/a	n/a	0ms	2.33x	n/a	n/a	1.00x
	(10000000,)	10000000	88ms	n/a	n/a	46ms	1.89x	n/a	n/a	1.00x
	(100, 100000)	10000000	95ms	n/a	n/a	9ms	10.07x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	9ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	78ms	n/a	n/a	n/a	1.00x
move_exp_nansum	(1000,)	1000	0ms	n/a	n/a	0ms	1.36x	n/a	n/a	1.00x
	(10000000,)	10000000	62ms	n/a	n/a	33ms	1.88x	n/a	n/a	1.00x
	(100, 100000)	10000000	71ms	n/a	n/a	7ms	9.70x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	6ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	60ms	n/a	n/a	n/a	1.00x
move_exp_nanvar	(1000,)	1000	0ms	n/a	n/a	0ms	1.40x	n/a	n/a	1.00x
	(10000000,)	10000000	77ms	n/a	n/a	42ms	1.82x	n/a	n/a	1.00x
	(100, 100000)	10000000	84ms	n/a	n/a	9ms	9.71x	n/a	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	n/a	9ms	n/a	n/a	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	n/a	73ms	n/a	n/a	n/a	1.00x
nanargmax1	(1000,)	1000	0ms	0ms	n/a	0ms	13.07x	0.21x	n/a	1.00x
	(10000000,)	10000000	31ms	12ms	n/a	12ms	2.45x	1.00x	n/a	1.00x
	(100, 100000)	10000000	28ms	13ms	n/a	13ms	2.16x	1.00x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	13ms	n/a	13ms	n/a	1.05x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	133ms	n/a	127ms	n/a	1.05x	n/a	1.00x
nanargmin1	(1000,)	1000	0ms	0ms	n/a	0ms	12.72x	0.21x	n/a	1.00x
	(10000000,)	10000000	27ms	13ms	n/a	12ms	2.19x	1.01x	n/a	1.00x
	(100, 100000)	10000000	26ms	13ms	n/a	12ms	2.05x	1.02x	n/a	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	13ms	n/a	13ms	n/a	1.05x	n/a	1.00x
	(100, 1000, 1000)	100000000	n/a	135ms	n/a	129ms	n/a	1.05x	n/a	1.00x
nancount	(1000,)	1000	0ms	n/a	0ms	0ms	2.24x	n/a	0.05x	1.00x
	(10000000,)	10000000	5ms	n/a	4ms	3ms	1.40x	n/a	1.06x	1.00x
	(100, 100000)	10000000	9ms	n/a	3ms	1ms	11.00x	n/a	4.16x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	4ms	1ms	n/a	n/a	3.58x	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	45ms	7ms	n/a	n/a	6.74x	1.00x
nanmax1	(1000,)	1000	0ms	0ms	0ms	0ms	8.21x	0.21x	0.38x	1.00x
	(10000000,)	10000000	41ms	12ms	1ms	13ms	3.26x	1.00x	0.11x	1.00x
	(100, 100000)	10000000	45ms	41ms	1ms	13ms	3.62x	3.24x	0.11x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	40ms	1ms	12ms	n/a	3.31x	0.12x	1.00x
	(100, 1000, 1000)	100000000	n/a	402ms	15ms	121ms	n/a	3.31x	0.12x	1.00x
nanmean	(1000,)	1000	0ms	0ms	0ms	0ms	1.32x	0.02x	0.20x	1.00x
	(10000000,)	10000000	23ms	9ms	27ms	10ms	2.42x	0.98x	2.83x	1.00x
	(100, 100000)	10000000	28ms	9ms	27ms	2ms	13.58x	4.54x	13.13x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	9ms	27ms	2ms	n/a	4.56x	13.69x	1.00x
	(100, 1000, 1000)	100000000	n/a	91ms	310ms	17ms	n/a	5.39x	18.39x	1.00x
nanmin1	(1000,)	1000	0ms	0ms	0ms	0ms	8.09x	0.21x	0.38x	1.00x
	(10000000,)	10000000	41ms	12ms	1ms	13ms	3.27x	1.00x	0.11x	1.00x
	(100, 100000)	10000000	45ms	41ms	1ms	13ms	3.62x	3.24x	0.11x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	40ms	1ms	12ms	n/a	3.28x	0.12x	1.00x
	(100, 1000, 1000)	100000000	n/a	401ms	15ms	122ms	n/a	3.30x	0.12x	1.00x
nanquantile	(1000,)	1000	0ms	n/a	0ms	0ms	1.46x	n/a	0.57x	1.00x
	(10000000,)	10000000	186ms	n/a	155ms	198ms	0.94x	n/a	0.78x	1.00x
	(100, 100000)	10000000	197ms	n/a	181ms	36ms	5.45x	n/a	5.01x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	n/a	425ms	34ms	n/a	n/a	12.50x	1.00x
	(100, 1000, 1000)	100000000	n/a	n/a	4254ms	331ms	n/a	n/a	12.85x	1.00x
nanstd	(1000,)	1000	0ms	0ms	0ms	0ms	1.06x	0.06x	0.46x	1.00x
	(10000000,)	10000000	29ms	29ms	53ms	19ms	1.50x	1.51x	2.75x	1.00x
	(100, 100000)	10000000	33ms	29ms	53ms	4ms	8.29x	7.35x	13.27x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	28ms	55ms	4ms	n/a	7.25x	14.43x	1.00x
	(100, 1000, 1000)	100000000	n/a	294ms	600ms	37ms	n/a	8.02x	16.35x	1.00x
nansum	(1000,)	1000	0ms	0ms	0ms	0ms	1.28x	0.02x	0.08x	1.00x
	(10000000,)	10000000	22ms	9ms	24ms	10ms	2.28x	0.97x	2.52x	1.00x
	(100, 100000)	10000000	27ms	9ms	24ms	2ms	17.71x	6.24x	16.05x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	9ms	25ms	1ms	n/a	6.05x	16.66x	1.00x
	(100, 1000, 1000)	100000000	n/a	90ms	282ms	13ms	n/a	6.71x	21.07x	1.00x
nanvar	(1000,)	1000	0ms	0ms	0ms	0ms	1.08x	0.06x	0.45x	1.00x
	(10000000,)	10000000	28ms	28ms	53ms	19ms	1.50x	1.49x	2.81x	1.00x
	(100, 100000)	10000000	33ms	28ms	54ms	4ms	8.18x	6.97x	13.32x	1.00x
	(10, 10, 10, 10, 1000)	10000000	n/a	28ms	56ms	4ms	n/a	7.13x	14.28x	1.00x
	(100, 1000, 1000)	100000000	n/a	281ms	601ms	32ms	n/a	8.71x	18.65x	1.00x

²³⁴⁵

Example implementation

Numbagg makes it easy to write, in pure Python/NumPy, flexible aggregation functions accelerated by Numba. All the hard work is done by Numba's JIT compiler and NumPy's gufunc machinery (as wrapped by Numba).

For example, here is how we wrote nansum:

import numpy as np
from numbagg.decorators import ndreduce

@ndreduce.wrap()
def nansum(a):
    asum = 0.0
    for ai in a.flat:
        if not np.isnan(ai):
            asum += ai
    return asum

Implementation details

Numbagg includes somewhat awkward workarounds for features missing from NumPy/Numba:

• It implements its own cache for functions wrapped by Numba's guvectorize, because that decorator is rather slow.
• It does its own handling of array transposes to handle the axis argument in reduction functions.
• It rewrites plain functions into gufuncs, to allow writing a traditional function while retaining the multidimensional advantages of gufuncs.

Already some of the ideas here have flowed upstream to numba (for example, an axis parameter), and we hope that others will follow.

License

3-clause BSD. Includes portions of Bottleneck, which is distributed under a Simplified BSD license.

Footnotes

1. This function is not currently parallelized, so exhibits worse performance on parallelizable arrays. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. Benchmarks were run on a Mac M1 laptop in December 2023 on numbagg's HEAD, pandas 2.1.1, bottleneck 1.3.7, numpy 1.25.2, with python numbagg/test/run_benchmarks.py -- --benchmark-max-time=10. They run in CI, though GHA's low CPU count means we don't see the full benefits of parallelization. ↩

3. While we separate the setup and the running of the functions, pandas still needs to do some work to create its result dataframe, and numbagg does some checks in python which bottleneck does in C or doesn't do. So use benchmarks on larger arrays for our summary so we can focus on the computational speed, which doesn't asymptote away. Any contributions to improve the benchmarks are welcome. ↩

4. In some instances, a library won't have the exact function — for example, pandas doesn't have an equivalent move_exp_nancount function, so we use its sum function on an array of 1s. Similarly for group_nansum_of_squares, we use two separate operations. ↩

5. anynan & allnan are also functions in numbagg, but not listed here as they require a different benchmark setup. ↩