rddsketch

Python bindings, via PyO3, for sketches-ddsketch — a Rust port of DataDog's Go/Java DDSketch implementations. A near-drop-in, faster replacement for DataDog's official ddsketch Python package.

Install

pip install rddsketch

Prebuilt wheels (Linux/macOS/Windows, Python 3.11+) via abi3, no Rust toolchain needed at install time.

Usage

from rddsketch import DDSketch

s = DDSketch(relative_accuracy=0.01)   # default: 1% relative error
s.add(12.3)
s.add(45.6)

s.quantile(0.5)                        # -> float | None
s.count, s.sum, s.min, s.max

s2 = DDSketch(relative_accuracy=0.01)
s2.add(100.0)
s.merge(s2)                            # in-place, requires matching config

s3 = s.copy()
data = s.to_bytes()
restored = DDSketch.from_bytes(data, relative_accuracy=0.01)

add/quantile/get_quantile_value/merge/count/sum/min/max match ddsketch.DDSketch directly. Differences: no weight param on add, constructor args and serialization format differ, avg is an added convenience property.

Benchmarks

Single-threaded, release build, rddsketch vs. ddsketch:

n	rddsketch	ddsketch	speedup
10,000	0.001s	0.012s	14.9x
100,000	0.009s	0.069s	7.4x
1,000,000	0.084s	0.642s	7.6x
10,000,000	0.795s	6.741s	8.5x

Quantile accuracy is identical (same algorithm). Memory usage is similar for a single sketch; for many independent sketches rddsketch uses meaningfully less. Full numbers and methodology: benchmarks/results.md.

Design notes

• Algorithm: sketches-ddsketch does the real work; this crate validates inputs and wraps it for Python.
• Bin store: dense, collapsing, bounded by max_num_bins (default 2048, matching LogCollapsingLowestDenseDDSketch) — not the unbounded ddsketch.DDSketch.
• Serialization: to_bytes()/from_bytes() use DataDog's Java-compatible binary encoding unmodified. from_bytes() needs relative_accuracy passed back in — it can't be recovered from the bytes alone.
• No weight param: the underlying store counts unweighted.