flypipe 5.0.0

Flypipe

Flypipe is a Python framework to simplify development, management and maintenance of transformation pipelines, which are commonly used in the data, feature and ML model space.

Each transformation is implemented in a small, composable function, a special decorator is then used to define it as a Flypipe node, which is the primary model Flypipe uses. Metadata on the node decorator allows for multiple nodes to be linked together into a Directed Acyclic Graph (DAG).

As each node (transformation) is connected to its ancestors, we can easily view the pipeline graphically in a html page (my_graph.html()) or execute it by invoking my_graph.run()

Installation

pip install flypipe

Package published in PyPI.

Example Usage

import pandas as pd
from flypipe.node import node


@node(
  type="pandas"
)
def t0():
  return pd.DataFrame(data={"fruit": ["mango", "lemon"]})


@node(
  type="pandas",
  dependencies=[t0.select("fruit").alias("df")]
)
def t1(df):
  categories = {'mango': 'sweet', 'lemon': 'sour'}
  df['flavour'] = df['fruit']
  df = df.replace({'flavour': categories})
  return df

df = t1.run()

>>> print(df)
+----+---------+-----------+
|    | fruit   | flavour   |
+====+=========+===========+
|  0 | mango   | sweet     |
+----+---------+-----------+
|  1 | lemon   | sour      |
+----+---------+-----------+

What Flypipe aims to facilitate?

• Free open-source tool for data transformations
• Facilitate streaming pipeline development (improved use of caches)
• Increase pipeline stability (better use of unittests)
• End-to-end transformation lineage
• Create development standards for Data Engineers, Machine Learning Engineers and Data Scientists
• Improve re-usability of transformations in different pipelines & contexts via composable nodes
• Faster integration and portability of pipelines to different contexts with different available technology stacks:
  • Flexibility to use and mix up pyspark/pandas on spark/pandas in transformations seamlessly
  • As a simple wheel package, it's very lightweight and unopinionated about runtime environment. This allows for it to be easily integrated into Databricks and independently of Databricks.
• Low latency for on-demand feature generation and predictions
• Framework level optimisations and dynamic transformations help to make even complex transformation pipelines low latency. This in turn allows for on-demand feature generation/predictions.

Commonly used

Source Code

API code is available at https://github.com/flypipe/flypipe.

Documentation

Full documentation is available at https://flypipe.github.io/flypipe/.