BUG: avoid float upcast when mixing signed/unsigned ints in isin #62608
Conversation
pandas/core/algorithms.py
Outdated
| # If the dtypes differ and either side is unsigned integer, | ||
| # prefer object dtype to avoid unsafe upcast to float64 that | ||
| # can lose precision for large 64-bit integers. |
There was a problem hiding this comment.
Does this change the performance when values and comps are both integer like and fit within an integer type?
There was a problem hiding this comment.
The fix takes the conservative approach of converting values to [object] when mixing signed and unsigned integer dtypes to ensure correctness. This preserves exact integer equality but may be slower for very large arrays compared to a numeric-only path.
This trade-off is favoring correctness over the rare case of very large arrays with mixed signed/unsigned ints
There was a problem hiding this comment.
Another approach for better performance can be to remove the earlier asymmetric object-conversion block and add a fast, safe numeric path that correctly handles signed/unsigned mixes without converting to object.
The idea is to use masked uint64 lookups to avoid float casts and preserve performance. I’ll place this fast-path after the comps_array extraction and before the common-type coercion, by mapping signed int64 and uint64 values into the wider unsigned space and performing hashtable lookups on uint64.
This will involve changes roughly around lines algorithms.py+6-16. I’ll also run the new tests afterward to verify behavior. Not sure if that breaks something, should I try?
2 participants
…ts added)
doc/source/whatsnew/vX.X.X.rstfile if fixing a bug or adding a new feature.Summary
Fix incorrect Series.isin results when comparing signed int64 values with uint64 values that are not equal. Previously, mixing signed and unsigned 64-bit integers could trigger a numeric common-type coercion to float64 which may lose precision and produce false positives. This change prevents that unsafe upcast by preferring an object-based comparison when signed and unsigned integer types are mixed.
Root cause
When isin attempted to find a common numeric dtype between comps (left side) and values (right side), mixing signed int64 with uint64 could lead to casting both sides to float64. Converting large 64-bit integers to float64 loses precision and can make two distinct integers compare as equal.
What I changed
[algorithms.py]:
Adjusted the condition used before converting values to an object array so that when dtypes differ and either side is an unsigned integer, [values] is converted to object (i.e., Python-level equality / hashtable lookup) instead of numeric coercion. This makes the mixed signed/unsigned decision symmetric and avoids unsafe float upcasts.
[test_isin.py]
Added test_isin_int64_vs_uint64_mismatch which reproduces the reported case and asserts the correct False result.