Provium

Provium helps you build processing workflows whose results explain where they came from. Store a result as an artifact, use that artifact as input to another step, and save the new outputs as artifacts of their own. Provium records those relationships automatically as your workflow runs.

Each processing step is represented by a versioned procedure. When a procedure reads existing artifacts and creates new ones, Provium links the outputs to the procedure and its inputs. That lineage travels with every result, including its full upstream history, so a final artifact can be traced back through every intermediate result and the procedures that produced them.

This keeps provenance out of your application logic: you work with inputs, perform the computation, and write outputs inside a procedure execution. Provium handles the dependency graph, integrity metadata, and lifecycle of those artifacts for you.

Features

• Typed readers and writers for application-specific binary formats
• Automatic input, output, and procedure lineage
• SHA-256 payload integrity checks
• Streaming, body-relative binary I/O
• Runtime artifact discovery through Python entry points
• Optional configuration snapshots, including Pydantic v2 models
• No required runtime dependencies

Installation

Provium requires Python 3.12 or newer.

python -m pip install provium

Quick start

Provium includes JsonArtifact for storing JSON-compatible values. This example records a collection of measurements and produces a summary:

from provium import JsonArtifact, Procedure

COLLECT = Procedure(name="collect", version="1")
SUMMARIZE = Procedure(name="summarize", version="1")

with COLLECT.execute():
    measurements = JsonArtifact.create("measurements.pa")
    measurements.write({"measurements": [12.5, 14.0, 13.5]})

with SUMMARIZE.execute():
    measurements = JsonArtifact.open("measurements.pa")
    payload = measurements.read()
    assert isinstance(payload, dict)
    values = payload["measurements"]
    assert isinstance(values, list)
    readings = [float(value) for value in values]

    summary = JsonArtifact.create("summary.pa")
    summary.write(
        {
            "count": len(readings),
            "minimum": min(readings),
            "maximum": max(readings),
            "average": round(sum(readings) / len(readings), 2),
        }
    )

summary.pa contains count, minimum, maximum, and average, together with the lineage of the measurements and the procedures that collected and summarized them. A rendered graph looks like this, with identities shortened for readability:

flowchart LR
    collect(["collect 1<br/>collect-execution"])
    measurements["provium.artifact.prefab.json.JsonArtifact<br/>measurements-id"]
    summarize(["summarize 1<br/>summarize-execution"])
    summary["provium.artifact.prefab.json.JsonArtifact<br/>summary-id"]

    collect --> measurements
    measurements --> summarize
    summarize --> summary

When each context exits successfully, Provium closes its handles and finalizes its output files. If a context exits with an exception, its pending outputs are not committed. Readers and writers are bound to their execution and cannot be used after its context exits.

JsonArtifact uses deterministic UTF-8 JSON encoding and supports null, booleans, finite numbers, strings, arrays, and objects with string keys.

Custom artifact types

For an application-specific binary format, define reader, writer, and artifact classes. Here is the same number workflow using signed 64-bit integers:

import struct

from provium import Artifact, ArtifactReader, ArtifactWriter

INTEGER = struct.Struct(">q")


class IntegerReader(ArtifactReader):
    def read(self) -> int:
        return INTEGER.unpack(self.body.read(INTEGER.size))[0]


class IntegerWriter(ArtifactWriter):
    def write(self, value: int) -> None:
        self.body.write(INTEGER.pack(value))


class IntegerArtifact(Artifact[IntegerReader, IntegerWriter]):
    reader = IntegerReader
    writer = IntegerWriter

Use the custom type just like the prefab JSON artifact:

from provium import session

from your_package.artifacts import IntegerArtifact

SOURCE = Procedure(name="source", version="1")
ADD = Procedure(name="add", version="1")

with SOURCE.execute():
    left = IntegerArtifact.create("left.pa")
    left.write(2)

    right = IntegerArtifact.create("right.pa")
    right.write(3)

with ADD.execute():
    left = IntegerArtifact.open("left.pa")
    right = IntegerArtifact.open("right.pa")
    total = IntegerArtifact.create("sum.pa")
    total.write(left.read() + right.read())

The workflow has the same lineage, now with application-specific integer artifacts. Identities are again shortened in the diagram:

flowchart LR
    source(["source 1<br/>source-execution"])
    left["your_package.artifacts.IntegerArtifact<br/>left-id"]
    right["your_package.artifacts.IntegerArtifact<br/>right-id"]
    add(["add 1<br/>add-execution"])
    total["your_package.artifacts.IntegerArtifact<br/>sum-id"]

    source --> left
    source --> right
    left --> add
    right --> add
    add --> total

Registration is optional. Without it, Provium stores the artifact class's full path, such as your_package.artifacts.IntegerArtifact, as its identifier. Typed calls such as IntegerArtifact.open() can read these artifacts directly.

Register the artifact when you want a stable custom identifier, aliases, or dynamic loading through provium.open_artifact():

from provium import ArtifactCatalog

from .artifacts import IntegerArtifact

catalog = ArtifactCatalog()
catalog.register("example.IntegerV1", IntegerArtifact)

Expose that catalog from pyproject.toml so Provium can discover it:

[project.entry-points."provium.catalogs"]
example = "your_package.catalog:catalog"

Inspecting provenance

Every reader exposes the artifact header and lineage:

from provium import Procedure

from your_package.artifacts import IntegerArtifact

with session():
    artifact = IntegerArtifact.open("sum.pa")
    print(artifact.read())
    print(artifact.identity)
    print(artifact.artifact_identifier)
    print(artifact.lineage.to_json())

Use provium.open_artifact() when the concrete type should be resolved from the identifier stored in the file rather than selected in advance.

Reusing artifacts across procedures

A session records every artifact opened within it, even after its reader is closed. Procedure executions inherit those recorded inputs and create a nested session for artifacts used only by that execution:

from provium import Procedure, session

PREDICT = Procedure(name="predict", version="1")

with session():
    model_reader = ModelArtifact.open("model.pa")
    model = load_model(model_reader)
    model_reader.close()

    for input_path, output_path in jobs:
        with PREDICT:
            data = DataArtifact.open(input_path)
            result = model.predict(data.read())
            ResultArtifact.create(output_path).write(result)

Each result depends on the shared model and its own data artifact. Nested generic sessions similarly inherit artifacts recorded by their ancestors.

Calling a procedure is shorthand for execute(), including configured executions: with PREDICT(config=settings): ....

Command-line tools

Inspect an artifact's generic metadata without loading its concrete artifact type:

provium inspect result.pa

Generate Mermaid or Graphviz source for an artifact's complete lineage:

provium graph --renderer mermaid result.pa lineage.mmd
provium graph --renderer graphviz result.pa lineage.dot

Image output supports SVG, PNG, and PDF and defaults to the Mermaid renderer:

provium graph result.pa lineage.svg
provium graph --renderer graphviz result.pa lineage.png

Mermaid image rendering requires the official mmdc executable. Graphviz rendering requires the optional Python package and Graphviz system package:

npm install --global @mermaid-js/mermaid-cli
python -m pip install 'provium[visualization]'

The output type is inferred from its extension. Library callers can use the functions in provium.tool to produce Mermaid or DOT source and to receive rendered images as bytes.

Development

Create a virtual environment and install the project with its test dependencies:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[test]'

Run the test suite:

pytest

This also runs Ruff linting and ruff format --check over src and test. The project requires 100% statement and branch coverage for the provium package.