https://github.com/datadog/sketches-py
Python implementations of the distributed quantile sketch algorithm DDSketch
https://github.com/datadog/sketches-py
ddsketch quantile
Last synced: 6 months ago
JSON representation
Python implementations of the distributed quantile sketch algorithm DDSketch
- Host: GitHub
- URL: https://github.com/datadog/sketches-py
- Owner: DataDog
- License: other
- Created: 2018-09-24T13:20:19.000Z (about 7 years ago)
- Default Branch: master
- Last Pushed: 2024-09-03T22:01:50.000Z (about 1 year ago)
- Last Synced: 2025-04-02T04:04:00.736Z (6 months ago)
- Topics: ddsketch, quantile
- Language: Python
- Homepage:
- Size: 167 KB
- Stars: 86
- Watchers: 25
- Forks: 18
- Open Issues: 4
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
Awesome Lists containing this project
README
# ddsketch
This repo contains the Python implementation of the distributed quantile sketch
algorithm DDSketch [1]. DDSketch has relative-error guarantees for any quantile
q in [0, 1]. That is if the true value of the qth-quantile is `x` then DDSketch
returns a value `y` such that `|x-y| / x < e` where `e` is the relative error
parameter. (The default here is set to 0.01.) DDSketch is also fully mergeable,
meaning that multiple sketches from distributed systems can be combined in a
central node.
Our default implementation, `DDSketch`, is guaranteed [1] to not grow too large
in size for any data that can be described by a distribution whose tails are
sub-exponential.
We also provide implementations (`LogCollapsingLowestDenseDDSketch` and
`LogCollapsingHighestDenseDDSketch`) where the q-quantile will be accurate up to
the specified relative error for q that is not too small (or large). Concretely,
the q-quantile will be accurate up to the specified relative error as long as it
belongs to one of the `m` bins kept by the sketch. If the data is time in
seconds, the default of `m = 2048` covers 80 microseconds to 1 year.
## Installation
To install this package, run `pip install ddsketch`, or clone the repo and run
`python setup.py install`. This package depends on `numpy` and `protobuf`. (The
protobuf dependency can be removed if it's not applicable.)
## Usage
```
from ddsketch import DDSketch
sketch = DDSketch()
```
Add values to the sketch
```
import numpy as np
values = np.random.normal(size=500)
for v in values:
sketch.add(v)
```
Find the quantiles of `values` to within the relative error.
```
quantiles = [sketch.get_quantile_value(q) for q in [0.5, 0.75, 0.9, 1]]
```
Merge another `DDSketch` into `sketch`.
```
another_sketch = DDSketch()
other_values = np.random.normal(size=500)
for v in other_values:
another_sketch.add(v)
sketch.merge(another_sketch)
```
The quantiles of `values` concatenated with `other_values` are still accurate to within the relative error.
## Development
To work on ddsketch a Python interpreter must be installed. It is recommended to use the provided development
container (requires [docker](https://www.docker.com/)) which includes all the required Python interpreters.
docker-compose run dev
Or, if developing outside of docker then it is recommended to use a virtual environment:
pip install virtualenv
virtualenv --python=3 .venv
source .venv/bin/activate
### Testing
To run the tests install `riot`:
pip install riot
Replace the Python version with the interpreter(s) available.
# Run tests with Python 3.9
riot run -p3.9 test
### Release notes
New features, bug fixes, deprecations and other breaking changes must have
release notes included.
To generate a release note for the change:
riot run reno new
Edit the generated file to include notes on the changes made in the commit/PR
and add commit it.
### Formatting
Format code with
riot run fmt
### Type-checking
Type checking is done with [mypy](http://mypy-lang.org/):
riot run mypy
### Type-checking
Lint the code with [flake8](https://flake8.pycqa.org/en/latest/):
riot run flake8
### Protobuf
The protobuf is stored in the go repository: https://github.com/DataDog/sketches-go/blob/master/ddsketch/pb/ddsketch.proto
Install the minimum required protoc and generate the Python code:
```sh
docker run -v $PWD:/code -it ubuntu:18.04 /bin/bash
apt update && apt install protobuf-compiler # default is 3.0.0
protoc --proto_path=ddsketch/pb/ --python_out=ddsketch/pb/ ddsketch/pb/ddsketch.proto
```
### Releasing
1. Generate the release notes and use [`pandoc`](https://pandoc.org/) to format
them for Github:
```bash
git checkout master && git pull
riot run -s reno report --no-show-source | pandoc -f rst -t gfm --wrap=none
```
Copy the output into a new release: https://github.com/DataDog/sketches-py/releases/new.
2. Enter a tag for the release (following [`semver`](https://semver.org)) (eg. `v1.1.3`, `v1.0.3`, `v1.2.0`).
3. Use the tag without the `v` as the title.
4. Save the release as a draft and pass the link to someone else to give a quick review.
5. If all looks good hit publish
## References
[1] Charles Masson and Jee E Rim and Homin K. Lee. DDSketch: A fast and fully-mergeable quantile sketch with relative-error guarantees. PVLDB, 12(12): 2195-2205, 2019. (The code referenced in the paper, including our implementation of the the Greenwald-Khanna (GK) algorithm, can be found at: https://github.com/DataDog/sketches-py/releases/tag/v0.1 )