Deployment

driftflow 2.0.0 is a typed library release and a static documentation site. The runtime has no network or service dependency; deploy the package and site independently.

Package build

python3 -m pip install -e ".[dev]"
PYTHON=.venv/bin/python ./scripts/ci.sh

The CI script builds the sdist and wheel and runs pip-audit with the editable package excluded from the vulnerability scan. Install the wheel in a clean environment before publishing it.

Library integration

Models, schedules, controllers, and solvers are explicit collaborators:

from driftflow import (
    CommonConfig,
    DvsAdaptiveController,
    DVSSampler,
    GraphState,
    GruMApproximation,
    LinearSchedule,
    get_dataset_config,
)

state = GraphState.from_lists(
    features=[[0.0, 0.0], [0.0, 0.0]],
    adjacency=[[0.0, 0.0], [0.0, 0.0]],
)
model = GruMApproximation(num_nodes=2, feature_dim=2, seed=7)
dataset = get_dataset_config("GruM", "QM9")
controller = DvsAdaptiveController(CommonConfig(), dataset, gamma=0.22)
sampler = DVSSampler(model, LinearSchedule(0.01, 0.5), controller)

result = sampler.sample(state, terminal_time=1.0, seed=42)
print(result.diagnostics.total_steps, result.diagnostics.final_time)

For a custom denoiser, implement DriftModel.evaluate and return a validated DriftState. The approximation models included in this repository are educational stand-ins, not trained production denoisers.

Container smoke test

FROM python:3.12-slim
WORKDIR /app
COPY . .
RUN python -m pip install --no-cache-dir .
CMD ["python", "examples/demo.py", "--model", "GruM", "--dataset", "QM9", "--solver", "Euler-Maruyama", "--use-approximation"]

Static site

cd site
npm ci
npm run build

The Pages workflow publishes site/dist and copies the generated index to 404.html when Astro has not produced one. The site is configured for the project path /driftflow.

Operational checks

Monitor SampleResult.diagnostics rather than parsing logs. In particular, track total steps, final time, timestep bounds, and failure counts by ConfigurationError, ShapeError, NumericalStabilityError, and SamplingLimitError. Set an application-level timeout around sampling and keep max_steps finite for every deployment.