README.md

Pipeline development framework, easy to experiment and compare different pipelines, quick to deploy to workflow orchestration tools

---

Most of the current workflow orchestrators focus on executing the already-developed pipelines in production. This library focuses on the pipeline development process. It aims to make it easy to develop pipeline, and once the user reaches a good pipeline, it aims to make it easy to export to other production-grade workflow orchestrators. Notable features:

• Manage pipeline experiment: store pipeline run outputs, compare pipelines & visualize.
• Support pipeline as code: allow complex customization.
• Support pipeline as configuration - suitable for plug-and-play when pipeline is more stable.
• Fast pipeline execution, auto-cache & run from cache when necessary.
• Allow version control of artifacts with git-lfs/dvc/god...
• Export pipeline to compatible workflow orchestration tools (e.g. Argo workflow, Airflow, Kubeflow...).

Install

pip install theflow

Quick start

(A code walk-through of this session is stored in examples/10-minutes-quick-start.ipynb. You can run it with Google Colab (TODO - attach the link).)

Pipeline can be defined as code. You initialize all the ops in self.initialize and route them in self.run.

from theflow import Function

# Define some operations used inside the pipeline
# Operation 1: normal class-based Python object
class IncrementBy(Function):

  x: int

  def run(self, y):
    return self.x + y

# Operation 2: normal Python function
def decrement_by_5(x):
  return x - 5

# Declare flow
class MathFlow(Function):

  increment: Function
  decrement: Function

  def run(self, x):
    # Route the operations in the flow
    y = self.increment(x)
    y = self.decrement(y)
    y = self.increment(y)
    return y

flow = MathFlow(increment=IncrementBy(x=10), decrement=decrement_by_5)

You run the pipeline by directly calling it. The output is the same object returned by self.run.

output = flow(x=5)
print(f"{output=}, {type(output)=}")      # output=5, type(output)=int

You can investigate pipeline's last run through the last_run property.

flow.last_run.id()                        # id of the last run
flow.last_run.logs()                      # list all information of each step
# [TODO] flow.last_run.visualize(path="vis.png")   # export the graph in `vis.png` file

Future features

• Arguments management
• Cache
  • cache by runs, organized by root task, allow reproducible
  • specify the files
  • the keys are like lru_cache, takes in the original input key, specify the cache, but the cache should be file-backed, for run-after-run execution.
  • cli command to manipulate cache
• Compare pipeline in a result folder
• Dynamically create reproducible config
• Support pipeline branching and merging
• Support single process or multi-processing pipeline running
• Can synchronize changes in the workflow, allowing logs from different run to be compatible with each other
• Compare different runs
  • Same cache directory
  • Compare evaluation result based on kwargs
• CLI List runs
• CLI Delete unnecessary runs
• Add coverage, pre-commit, CI...

License

MIT License.