[paddymul]

This PR fixes the xorq side (#1055) and, since #1072 was merged into this branch, the pandas and polars side. #1072 went in with only its test commit (b0c69d3); its fix commit is ac5a0cd here.

Problem

xorq: _numeric_histogram in xorq_stats_v2.py buckets with (x - min) / (max - min) * 10 cast to int64. A float column holding inf or -inf makes max - min infinite, so that row's bucket is NaN and the query fails with Can't cast value NaN to type Int64. histogram's default=[] swallows the error (#687), so the column shows no histogram. histogram_bins computes [nan, inf, inf, ...] from the same infinite range.

pandas and polars: histogram_series and pl_histogram_series take the 1% and 99% tails with np.quantile / Series.quantile over the raw column, then histogram the values between them. With an infinity in the column, numpy's interpolation can return NaN (inf - inf, or inf * 0, inside np.quantile). Every comparison against a NaN tail is False, so the middle is empty, histogram_args comes back empty, and the column falls to the categorical histogram (top value counts) with no histogram_bins, so color_map has nothing to read. Where the quantile is inf instead, the middle includes every finite value, the finite buckets stretch to the maximum, and the tail label reads infT–infT. Which one happens depends on where 0.99 * (n - 1) falls beside the infinity: with a single +inf, the quantile is NaN at 50 and 101 non-null values and inf at 99 and 100.

Change

xorq: when a float column's min or max is infinite, _numeric_histogram and histogram_bins call _finite_range, which runs one query for the min and max of the column's finite values. The histogram's bucket query drops the infinite rows, so the populations sum to 100% of the finite values, and histogram_bins spans the finite range. min and max still report the infinities. A column with no finite values, or one, still gets no histogram. A column with a finite range runs no extra query. The issue allowed either leaving the infinities out of the buckets or counting them in the end buckets. I left them out: an end bucket that includes inf would be labelled with a finite range it doesn't cover. histogram_bins now takes expr, execute and orig_col_name for that query, so the pipeline treats it as a query stat (_is_query_func) and caches its cell the way it caches histogram's.

pandas and polars: both functions take the tails over the finite values. The infinities fall outside them and land in the tail buckets, and the buckets of the finite values are the same with or without the infinities. The polars filter applies to float dtypes only, so integer and Decimal columns take the path they did before. A numeric column with 5 or fewer distinct values is still categorical (an infinity counts as a distinct value), and a column with no finite values still gets the categorical histogram.

fmt_num returns inf, -inf or nan for a non-finite bound, so a tail reads 198–inf and -inf<>1.5 instead of infT.

df_with_infinity no longer fails the warm-load check in test_stat_cache.py: the strict #1057 xfail on it XPASSed, so its entry comes out of _DDD_KNOWN_FAILURES, as that table's comment says to do. #1057 still applies to pl_df_with_temporal_edges.

New DDD frames: df_with_infinities_among_values and pl_df_with_infinities_among_values (101-row float columns where np.quantile is NaN, plus a control without infinities; the polars frame adds a real NaN and a null). They are on the docs DDD page, and test_stat_cache.py picks them up automatically.

Tests

xorq, TestHistogram in tests/unit/test_xorq_stats_v2.py, each for +inf, -inf and both: the histogram is non-empty and sums to 100% with min/max still reporting the infinities, and histogram_bins has 11 finite edges spanning the finite values. These failed on CI before the xorq fix (run 37506145362, Python / Test (Max Versions) (3.13): 6 failed, 1508 passed).

pandas (tests/unit/histogram_test.py) and polars (TestPlHistogram in tests/unit/test_pl_stats_v2.py):

• a sweep over 5 infinity shapes (+inf, -inf, two -inf, both, 5 on each side) and 12 finite-value counts that straddle the NaN/inf boundary; polars also runs each alone, beside a NaN, and beside a null. Each asserts a numeric histogram (no categorical entries), 11 finite histogram_bins, and the same finite buckets as the column without the infinities
• every DDD float column that holds an infinity beside more than 5 distinct finite values (ddd_float_columns_with_infinities in tests/unit/test_utils.py reads them from ddd_library, so frames added later are covered); the xorq test file runs the same columns
• pandas float64, float32 and nullable Float64
• tail buckets name the infinity, and fmt_tail_bucket formats infinite bounds as inf

These failed on CI before the pandas/polars fix (run 37513631803 on #1072's branch, and run 37513943690 on this PR after #1072's test commit merged: Python / Test (3.14), 240 failed, 997 passed; all 240 are the new tests). The fix commit adds guards that pass either way: columns with nothing to bucket, polars Float32, and the xorq DDD check. Locally the full unit suite passes on this branch (1940 passed, 17 skipped, 15 xfailed).

Not in this PR

Found while checking; each is separate from these fixes:

• a nullable Float64 column holding an infinity raises in mean, median and std (pandas stats: a nullable Float64/Int64 column loses mean, median and std when its std is pd.NA #1074)
• _formatted_char_count (customizations/styling.py:25) guards NaN but not infinity, so int(max_abs) raises OverflowError and the column's styling falls back to the parent class with a logged warning. df_with_infinity does this too. (styling: a float column holding an infinity falls back to the obj displayer (OverflowError in _formatted_char_count) #1075)
• xorq: a float column holding a real NaN (not a null) gets an empty histogram and [nan, ...] bin edges, because max is NaN (xorq stats: a float column holding a NaN gets no histogram and all-NaN histogram_bins #1076)

Closes #1055.

🤖 Generated with Claude Code

[github-actions]

📦 TestPyPI package published

pip install --index-strategy unsafe-best-match --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ buckaroo==0.15.9.dev37516214008

or with uv:

uv pip install --index-strategy unsafe-best-match --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ buckaroo==0.15.9.dev37516214008

MCP server for Claude Code

claude mcp add buckaroo-table -- uvx --from "buckaroo[mcp]==0.15.9.dev37516214008" --index-strategy unsafe-best-match --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ buckaroo-table

📖 Docs preview

🎨 Storybook preview

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.