fix: replace hardcoded fill value with dynamic min_nonzero/10 in get_binned_data

[mukund1985]

Summary

Fixes #334

The zero-fill logic in get_binned_data used a hardcoded threshold and fallback of 0.0001. When all non-zero percentages in a distribution are smaller than 0.0001 (common with large datasets or rare categories), the fill value becomes larger than legitimate data values. This makes KL-divergence and other stattest calculations incorrect — the fill is supposed to be a negligible epsilon, not a dominant value.

Root cause:

# BEFORE — fill can exceed real values when min_nonzero <= 0.0001
np.place(reference_percents, reference_percents == 0,
    min(reference_percents[reference_percents != 0]) / 10**6
    if min(reference_percents[reference_percents != 0]) <= 0.0001
    else 0.0001)

Fix:

# AFTER — always proportional to the smallest real value in that array
ref_nonzero = reference_percents[reference_percents != 0]
if len(ref_nonzero) > 0:
    np.place(reference_percents, reference_percents == 0, min(ref_nonzero) / 10)

min_nonzero / 10 is guaranteed strictly smaller than any real data value at any scale. The empty-array guard prevents errors when one side has no non-zero entries.

Changes

• src/evidently/legacy/calculations/stattests/utils.py — pandas implementation
• src/evidently/legacy/spark/calculations/stattests/utils.py — Spark implementation

Both use identical logic with their respective parameter names (feel_zeroes / fill_zeroes).

Test plan

• All 156 existing stattest unit tests pass (pytest tests/multitest/metrics/test_data_drift.py tests/stattests/ -v)
• Smoke test confirms fill value equals exactly min_nonzero / 10 and old hardcoded 0.0001 is gone
• Verified fix handles edge case where all values in one array are non-zero (empty guard)

[mukund1985]

Hey, just flagging — ran the existing test suite locally and everything passes. The fix is pretty minimal, just handling that edge case. Happy to change the approach if there's a better way to do it, just let me know.

[mukund1985]

@Liraim — would appreciate a review when you get a chance. Tests pass locally, happy to make any changes needed.

[mukund1985]

@DimaAmega — looks like CI hasn't triggered yet, could you approve the workflow run when you get a chance?

[github-actions]

📚 Artifacts deployed to GitHub Pages: https://evidentlyai.github.io/evidently/ci/#pr-1862-fix-fill-zeroes-dynamic-value