Working With Notebooks

Working with Notebooks

How to efficiently debug, fix, and validate notebooks in pymc-marketing.

1. The Mock Runner (fast local testing)

The mock runner injects mocked PyMC sampling into notebooks, letting you check for runtime errors without real MCMC. See scripts/run_notebooks/README.md.

Run a single notebook locally

uv run python scripts/run_notebooks/runner.py --notebooks docs/source/notebooks/mmm/mmm_example.ipynb

Run multiple notebooks

uv run python scripts/run_notebooks/runner.py --notebooks docs/source/notebooks/clv/sbg.ipynb docs/source/notebooks/mmm/mmm_example.ipynb

Run by category (CI-like)

uv run python scripts/run_notebooks/runner.py --exclude-dirs mmm    # non-MMM only
uv run python scripts/run_notebooks/runner.py --exclude-dirs clv bass customer_choice general  # MMM only

2. Real Notebook Execution (for functional code changes)

When source code changes affect model outputs, notebooks must be re-executed with real sampling to generate updated output cells.

Recommended: jupyter execute --inplace

# Clear stale outputs first
uv run jupyter nbconvert --clear-output docs/source/notebooks/mmm/mmm_example.ipynb

# Real execution with full MCMC
uv run jupyter execute docs/source/notebooks/mmm/mmm_example.ipynb --inplace

# Clean up auto-generated checkpoint dirs
find docs/source/notebooks/mmm -name .ipynb_checkpoints -type d -exec rm -rf {} +

Alternative: jupyter nbconvert

jupyter nbconvert --to notebook --execute --inplace docs/source/notebooks/mmm/mmm_example.ipynb

When to execute vs mock

3. Core Workflow (Iteration Loop)

CRITICAL: Never push until all targeted notebooks pass locally.

1. Identify failing notebooks (from CI or initial mock run)
2. FOR EACH notebook (one at a time):
   a. Run mock runner: python scripts/run_notebooks/runner.py --notebooks <path>
   b. Read the error (Cell In[X], exception type, traceback)
   c. Identify the pattern (see §4 below)
   d. Fix the notebook cell
   e. Repeat a-d until mock runner passes
3. Run ruff on changed files:
   ruff check --fix <files>
   ruff format <files>
4. Run pre-commit (or at minimum: ruff check + ruff format + mypy)
5. Push ONCE when all notebooks pass

4. Common Notebook Failure Patterns

B10: Bare DataArray/ndarray passed to plot_dist/plot_forest

# ❌ BROKEN — KeyError: 'posterior' or IndexError
azp.plot_dist(my_dataarray)
azp.plot_dist(my_numpy_array)

# ✅ FIXED — wrap in Dataset
azp.plot_dist(my_dataarray.to_dataset(name="var_name"))
azp.plot_dist(xr.Dataset({"x": xr.DataArray(data, dims=["sample"])}))

B11: figsize → figure_kwargs

# ❌ — ValueError: extra keywords
azp.plot_forest(idata, figsize=(8, 8))

# ✅
azp.plot_forest(idata, figure_kwargs={"figsize": (8, 8)})

B12: rug=True removed

# ❌ — ValueError: extra keywords
azp.plot_dist(idata, rug=True)
# ✅ — just remove rug=True
azp.plot_dist(idata)

B13: sample_dims mismatch

# ❌ — ValueError: sample_dims not found
azp.plot_dist(xr.Dataset({"x": data}), sample_dims=["sample"])  # data has chain/draw

# ✅ — remove sample_dims (auto-detect) or match data dims
azp.plot_dist(xr.Dataset({"x": data}))  # auto-detect for chain/draw data
azp.plot_dist(xr.Dataset({"x": data}), sample_dims=["sample"])  # for pre-stacked data

B14: List-based plot_forest with model_names

# ❌ — TypeError: list indices must be integers
azp.plot_forest([idata1, idata2], model_names=["A", "B"], ...)

# ✅ — combine into single DataTree with model dim
combined = xr.concat([ds1, ds2], dim="model").assign_coords(model=["A", "B"])
dt = xr.DataTree.from_dict({"/posterior": xr.DataTree(combined)})
azp.plot_forest(dt, ...)

B15/B16: Return type — PlotCollection, not axes

# ❌ — TypeError: 'PlotCollection' object is not subscriptable
ax = azp.plot_forest(idata, ...); ax[0].set_title("...")

# ✅ — extract axes from PlotCollection
pc = azp.plot_forest(idata, ...)
ax = pc.viz["/"]["figure"].values.item().axes[0]
ax.set_title("...")

B17: compact=True removed

# ❌ — ValueError: extra keywords
azp.plot_trace(idata, compact=True)
# ✅ — remove
azp.plot_trace(idata)

B18: group parameter for posterior_predictive

# ❌ — KeyError: 'posterior'
azp.plot_dist(pp_idata)  # pm.sample_posterior_predictive result
# ✅
azp.plot_dist(pp_idata, group="posterior_predictive")

B20: Missing xarray import

# ❌ — NameError: name 'xr' is not defined
# ✅ — add to imports cell:
import xarray as xr

B22: hdi_prob → prob, ci_bounds → ci_bound

# ❌
az.hdi(data, hdi_prob=0.94)
hdi.sel(ci_bounds="lower"), hdi.sel(ci_bounds="higher")
# ✅
az.hdi(data, prob=0.94)
hdi.sel(ci_bound="lower"), hdi.sel(ci_bound="upper")

B23: HDI integer indexing [0]/[1] → .sel(ci_bound=...)

# ❌ — IndexError
hdi_result[0], hdi_result[1]
# ✅
hdi_result.sel(ci_bound="lower"), hdi_result.sel(ci_bound="upper")

B24: az.plot_density removed

# ❌
az.plot_density(data, ...)
# ✅
azp.plot_dist(data)

B25: prob= in plot_dist/plot_forest (not same as ci_prob=)

# ❌ — ValueError: extra keywords
azp.plot_dist(idata, prob=1)
# ✅ — prob= is NOT a valid kwarg for plot_dist/plot_forest
azp.plot_dist(idata)  # just remove prob=

5. Identifying Failing Notebooks from CI

The CI runs notebooks in batches on the Test Notebooks workflow:

1. Go to the PR checks tab on GitHub
2. Find entries like notebooks (cvm, --notebooks clv) with ❌
3. Click through to the job log
4. Search for Error running notebook: — this gives the notebook name
5. Search for Exception encountered at "In [X]": — this gives the cell number

Alternatively, find the run ID with gh pr checks or gh run list:

# List recent notebook CI runs
gh run list --branch chore/pymc6-migrate --workflow "Test Notebooks" --limit 5

# Get error details from a specific run
gh run view <run_id> --log --job <job_id> | grep -A10 'Exception encountered'

6. Pre-Push Checklist

Before pushing notebook changes:

1. ☐ Mock runner passes for ALL changed notebooks
2. ☐ If source code changed: real execution passes for affected notebooks
3. ☐ ruff check --fix <files> && ruff format <files> on changed files
4. ☐ ruff check docs/source/notebooks/ --exclude 'docs/source/notebooks/*/dev/*' passes
5. ☐ ruff format --check docs/source/notebooks/ shows no reformats needed
6. ☐ pre-commit run --all-files passes locally
7. ☐ Push ONCE

7. Real Execution for Release

When re-executing notebooks for a release, use this full workflow per notebook. See notebook-release-changes.md for version-specific dependency bumps and API migrations.

Per-Notebook Workflow

# 1. Pre-flight smoke test (catches code errors fast)
uv run python scripts/run_notebooks/runner.py --notebooks docs/source/notebooks/category/my_notebook.ipynb

# 2. Clear stale outputs
uv run jupyter nbconvert --clear-output docs/source/notebooks/category/my_notebook.ipynb

# 3. Real execution with full MCMC sampling
uv run jupyter execute docs/source/notebooks/category/my_notebook.ipynb --inplace

# 4. Clean up auto-generated checkpoint dirs
find docs/source/notebooks/category -name .ipynb_checkpoints -type d -exec rm -rf {} +

# 5. Validate (check for errors + pre-commit)
uv run python -c "
import json
with open('docs/source/notebooks/category/my_notebook.ipynb') as f:
    nb = json.load(f)
for i, cell in enumerate(nb['cells']):
    for out in cell.get('outputs', []):
        if out.get('output_type') == 'error':
            print(f'ERROR in cell {i}: {out[\"evalue\"]}')
"

uv run pre-commit run --files docs/source/notebooks/category/my_notebook.ipynb

Fast Figure Iteration (Avoid Slow Notebook Re-Executions)

Re-executing a full notebook just to tweak a plot is painfully slow (MCMC takes minutes). Instead, isolate the figure code into a standalone script in sandbox/:

1. Extract the figure cell into sandbox/test_something.py
2. Mock data with the same shape/dims as the real idata (use rng = np.random.default_rng(42) for reproducibility)
3. Run the script — no MCMC, no data loading, just the plot code
4. Iterate with the user until the output looks right
5. Apply the final pattern back to the notebook

# 5-second iteration loop vs 10-minute notebook re-execution
uv run python sandbox/test_figure.py

Tmp outputs in sandbox/ are just for feedback, not kept. The point is speed: hours of notebook re-execution becomes seconds of isolated iteration. Used for plot_forest (delaxes, labels) and plot_dist (CI removal, legends).