Skip to content

stagedml/pylightnix

Repository files navigation

coverage

Pylightnix

Pylightnix is a minimalistic Python library for managing filesystem-based immutable data, inspired by Purely Functional Software Deployment Model thesis by Eelco Dolstra and the Nix package manager.

The library may be thought as of low-level API for creating caching wrappers for a subset of Python functions. In particular, Pylightnix allows user to

  • Prepare (or, in our terms - instantiate) the computation plan aimed at creating a tree of linked immutable stage objects, stored in the filesystem.
  • Implement (in our terms - realize) the prepared computation plan, access the resulting artifacts. Pylightnix is able to handle non-deterministic results of the computation.
  • Handle changes in the computation plan, re-use the existing artifacts whenever possible.
  • Gain full control over the cached data by inspecting the storage
import numpy as np
from pylightnix import *

fsinit('_pylightnix')

@autostage
def stage_dataset(A=100, file=[selfref, 'dataset.npy']):
  np.save(file, np.arange(0, A))

@autostage(nouts=10)
def stage_train(rindex:int, dataset=stage_dataset, model=[selfref, 'model.npy']):
  ds = np.load(dataset.file.val)
  np.save(model, rindex)

rrefs=realizeMany(instantiate(stage_train))
assert len(rrefs)>=10

Documentation

QuickStart [PDF] | API Referece [MD]

Demos:

  • GNU Hello, turn Pylightnix into a toy package manager to build the GNU Hello from sources.
  • MDRUN, evaluate code sections of a Markdown document, cache the results between runs.