Skip to content

System context

C4 level 1. badgers is a single Python library — there is no container level, no service boundary and no deployment topology to describe. What matters is who calls it and what it pulls in.

flowchart LR
    user["Data scientist /<br/>ML engineer"]
    pipeline["Evaluation or<br/>robustness pipeline"]

    badgers["badgers<br/>[Python library]<br/>generate(X) → Xt"]

    np["NumPy · pandas · SciPy<br/>[arrays, RNG, stats]"]
    sk["scikit-learn<br/>[estimators]"]
    nx["NetworkX<br/>[graphs]"]

    pymc["PyMC · PyTensor<br/>[optional extra: pymc]"]
    dowhy["DoWhy GCM<br/>[optional extra: dowhy]"]

    user -->|"writes generators,<br/>selectors, DGPs"| badgers
    pipeline -->|"imports badgers"| badgers
    badgers --> np
    badgers --> sk
    badgers --> nx
    badgers -.->|"lazy import"| pymc
    badgers -.->|"lazy import"| dowhy

Actors

Actor Interaction
Data scientist / ML engineer Uses the shipped generators to inject deficits into a dataset, then measures the effect on a model or analysis.
Generator author Subclasses GeneratorMixin and implements generate. See Creating Tabular Generators.
Causal modeller Supplies a fitted DGP (or wraps a DoWhy/PyMC model) and uses the causal generators. See Using DGPs.

External systems

System Kind Why it is there
NumPy hard dependency All data is handled as numpy arrays internally; the RNG is always caller-supplied.
pandas hard dependency Accepted as input and returned as output for tabular generators.
SciPy hard dependency Distributions used by noise and outlier samplers.
scikit-learn hard dependency Estimators used by drift detection and by the default DGP backend.
NetworkX hard dependency Graph structure for the causal generators and the graph missingness generator.
PyMC / PyTensor optional extra pymc One DGP backend (PyMCSCMDGP).
DoWhy GCM optional extra dowhy One DGP backend (GcmSCMDGP).

The two optional backends are imported lazily, inside their own modules, via a PEP 562 __getattr__ in badgers.core.dgp.backends. Importing badgers or badgers.core.dgp must never pull them in — this is asserted by tests/core/test_import_hygiene.py.

What badgers is not

  • Not a data-quality detector. It injects deficits; it does not find them.
  • Not a scikit-learn transformer. badgers.core.pipeline.Pipeline is deliberately not compatible with sklearn.pipeline.Pipeline — it chains generators, not estimators.
  • Not a data generator from scratch. It augments existing data; the input X is always required.