Python Polars 1.28.0

🚀 Performance improvements

• Lower Expr.(n_)unique to group_by on streaming engine (#22420)
• Chunk huge munmap calls (#22414)
• Add single-key variants of streaming group_by (#22409)
• Improve accumulate_dataframes_vertical performance (#22399)
• Use optimize rolling_quantile with varying window sizes (#22353)
• Dedicated rolling_skew kernel (#22333)
• Call large munmap's in background thread (#22329)
• New streaming group_by implementation (#22285)
• Patch jemalloc to not purge huge allocs eagerly if we have background threads (#22318)
• Turn on parallel=prefiltered by default for new streaming (#22190)

✨ Enhancements

• When reporting unexpected types in errors, module-qualify the typename (#22390)
• Add Series backward_fill / forward_fill (#22360)
• Add GPU support to sink_* APIs (#20940)
• Changed mapping type from dict to Mapping (#19400) (#19436)
• Make streaming dispatch public (#22347)
• Add rolling_kurtosis (#22335)
• Support Cast in IO plugin predicates (#22317)
• Add .sort(nulls_last=True) to booleans, categoricals and enums (#22300)
• Add rolling min/max for temporals (#22271)
• Support literal:list agg (#22249)
• Support running Polars SQL queries against any objects implementing the PyCapsule interface (#22235)
• Support implode + agg (#22230)
• Dispatch scans to new-streaming by default (#22153)

🐞 Bug fixes

• Ensure write_excel correctly preserves null values in nested dtype data on export (#22379)
• Panic when visualizing streaming physical plan with joins (#22404)
• Fix incorrect filter after LazyFrame.rename().select() (#22380)
• Fix select(len()) performance regression (#22363)
• Handle pytz named timezone in lit (#21785)
• Don't leak state during prefill CSE cache (#22341)
• Maintain float32 type in partitioned group-by (#22340)
• Resolve streaming panic on multiple merge_sorted (#22205)
• Fix ndjson nested types (#22325)
• Fix nested datetypes in ndjson (#22321)
• Check matching lengths for pl.corr (#22305)
• Move type coercion for pl.duration to planner (#22304)
• Check dtype to avoid panic with mixed types in min/max_horizontal (#21857)
• Coalesce correct column for new streaming full join (#22301)
• Don't collect NaN from Parquet Statistics (#22294)
• Set revmap for empty AnyValue to Series (#22293)
• Add an __all__ entry to internal type definition module (#22254)
• Datetime parser was incorrectly parsing 8-digit fractional seconds when format specified to expect 9 (#22180)
• More robust str → date conversion when reading from spreadsheet (#22276)
• Deprecate using is_in with 2 equal types and mark as elementwise (#22178)
• Duplicate key column name in streaming group_by due to CSE (#22280)
• Raise ColumnNotFoundError for missing columns in join_where (#22268)
• Parquet filters for logical types and operations (#22253)
• Ensure floating-point accuracy in hist (#22245)
• Check matching key datatypes for new streaming joins (#22247)
• Incorrect length BinaryArray/ListBuilder (#22227)

📖 Documentation

• Update docs for schema arg in scan_csv to match read_csv (#22357)
• Update pl.when documentation (#22345)
• Add missing is_business_day to documentation reference (#22338)
• Improve interpolation documentation to clarify behavior of null values (#22274)

🛠️ Other improvements

• Install pytorch for 3.13 on Windows (#22356)
• Make interpolate fix more robust (#22421)
• Fix interpolate test (#22417)
• Reduce hot table size in debug mode (#22400)
• Replace intrinsic with non-intrinsic (#22401)
• Make streaming dispatch public (#22347)
• Update rustc to 'nightly-2025-04-19' (#22342)
• Update mozilla-actions/sccache-action (#22319)
• Purge old parquet and scan code (#22226)
• Add an __all__ entry to internal type definition module (#22254)
• Add online skew/kurtosis algorithm for future use in rolling kernels (#22261)
• Add Polars Cloud 0.0.7 release notes (#22223)
• Change format name from list to implode (#22240)
• Make other parallel parquet modes filter afterwards (#22228)
• Close async reader issues (#22224)
• Add BinaryArrayBuilder (#22225)

Thank you to all our contributors for making this release possible!
@DavideCanton, @JakubValtar, @Jesse-Bakker, @MarcoGorelli, @NeejWeej, @Shoeboxam, @adamreeve, @alexander-beedie, @axellpadilla, @cmdlineluser, @coastalwhite, @d-reynol, @dongchao-1, @florian-klein, @kdn36, @math-hiyoko, @mcrumiller, @mroeschke, @nameexhaustion, @orlp, @ritchie46, @stijnherfst and @yiteng-guo