pmprov

Analytic provenance tracking middleware for exploratory process mining notebooks.

pmprov automatically records every meaningful decision an analyst makes during a Jupyter or Marimo session — which functions were called, with what arguments, on which data, and how the data changed — without requiring any modification to analyst code.

How it works

Your cell code  →  AST rewriter  →  trace_step()  →  DuckDB/SQLite + Parquet
                   (transparent)     (non-blocking)

Every top-level function call is intercepted at the AST level, wrapped with a lightweight tracer, and recorded as a node in a provenance tree. The analyst's variables contain exactly the same values they would without the middleware.

Getting started

Prerequisites: Python ≥ 3.13, uv

git clone https://github.com/alvarobernal2412/pmprov.git
cd pmprov
uv pip install -e ".[full]"

In a Jupyter notebook:

from tracker import init_jupyter, operation_type
import pandas as pd
import pm4py

# Register semantic labels for third-party functions (optional but recommended)
operation_type("data_loading",      pd.read_csv)
operation_type("process_discovery", pm4py.discover_petri_net_inductive)

# Start tracking — one call, no other changes to your code
rt = init_jupyter(history_name="My analysis")

# Your normal analysis code below — provenance is captured automatically
event_log = pd.read_csv("data/log.csv", parse_dates=["time:timestamp"])
net, im, fm = pm4py.discover_petri_net_inductive(event_log)

Resuming a history: calling init_jupyter/init_marimo again with the same history_name (and the same db_path) resumes the prior history by default — your kernel restarted, but the analysis lineage picks up where it left off. Pass fresh=True to force a brand-new history instead. Calling with no history_name always starts fresh, unaffected by this. Note this isn't safe across two kernels running concurrently against the same db_path + history_name — they'd race on whose writes "win".

In a Marimo notebook:

Marimo support requires a one-time setup step. The middleware patches Marimo's AST compiler at Python startup via sitecustomize.py, which must be placed in your virtual environment (already in place if installed via uv).

Structural requirement: Marimo's reactive DAG is built from the original cell code before the AST rewriter runs. This means the rewriter cannot inject dependency edges automatically. To guarantee that init_marimo() runs before any tracked cell, combine your path setup and init_marimo() call in one cell and export all imports (pd, constants, etc.) from it. Downstream cells then depend on those names naturally, which implicitly orders them after init — no need to add rt to every cell's parameters.

# ── Setup + init (one cell) ───────────────────────────────────────────────────
@app.cell
def _():
    import sys, pandas as pd
    from pathlib import Path
    from tracker import init_marimo, operation_type

    # ... path setup ...

    rt = init_marimo(history_name="My analysis")
    operation_type("data_loading", pd.read_csv)

    def settle():
        rt.storage._executor.submit(lambda: None).result()

    return pd, rt, settle, DATA_FILE  # export everything downstream cells need


# ── Analysis cells — no `rt` needed in params ─────────────────────────────────
@app.cell
def _(DATA_FILE, pd):          # depends on pd + DATA_FILE → init already ran
    event_log = pd.read_csv(DATA_FILE, parse_dates=["time:timestamp"])
    return (event_log,)

Run with:

uv run marimo run examples/marimo/provenance_demo.py
# or for interactive editing:
uv run marimo edit examples/marimo/provenance_demo.py

Note: Provenance capture works in Marimo. Some introspection features (interactive widget, manual checkout) are not yet validated in the Marimo environment.

Inspect the provenance graph:

rt.list_states()           # table of all recorded steps
rt.show_graph()            # static matplotlib tree
rt.show_graph_widget()     # interactive Jupyter widget (Jupyter only)

Branch from a past state:

branch = rt.checkout("paste-state-id-here", branch_name="experiment-v2")

See examples/jupyter/provenance_demo.ipynb for a full Jupyter walkthrough and examples/marimo/provenance_demo.py for the Marimo equivalent, both using the RTFM dataset.

Optional extras

uv pip install -e ".[full]"

Project structure

models/    Pydantic v2 domain models (provenance entities)
tracker/   Runtime middleware (AST rewriter, tracer, storage)
examples/  Demo notebook + RTFM dataset helpers

Running tests

uv sync --all-extras --dev
uv run pytest

Git hooks (optional)

Enforces the test suite and conventional commit format (≤ 45 chars) on every commit:

uv run pre-commit install && uv run pre-commit install --hook-type commit-msg

Skip this step if you prefer to run checks manually.

Contributing

See CONTRIBUTING.md.

License

Apache 2.0 — see LICENSE.