v26.08.00

What's Changed

🚨 Breaking Changes

• Enforce pandas 3.0 deprecations and subsequent related cuDF deprecations part 1 by @mroeschke in #20823
• Enforce Series.interpolate deprecations by @mroeschke in #20905
• Remove FutureWarning checks from series/test_pct_change by @mroeschke in #20904
• Enforce Series.replace argument deprecations by @mroeschke in #20903
• Enforce pandas 3 DataFrame/Series.first/last removal by @mroeschke in #20909
• Enforce pandas 3 fillna deprecation by @mroeschke in #20912
• Enforce pandas 3 timedelta frequency string deprecation by @mroeschke in #20915
• Enforce pandas 3 concat with empties deprecation by @mroeschke in #20916
• Enforce pandas 3 to_numeric(errors='ignore') deperecation by @mroeschke in #20920
• Enforce pandas 3 date_range(periods=) deprecation by @mroeschke in #20921
• Enforce pandas 3 resample deprecations by @mroeschke in #20923
• Enforce pandas 3 TimedeltaIndex/DatetimeIndex deprecations by @mroeschke in #20924
• Enforce pandas 3 future_stack argument switch by @mroeschke in #20926
• Remove pandas 3.0 Series.getitem int positional deprecation by @mroeschke in #20930
• Test pandas 3.0.0rc1 by @mroeschke in #20941
• Change to_datetime to match pandas-3 by @galipremsagar in #20913
• More misc cuDF classic test fixes for pandas 3 by @mroeschke in #21049
• Switch default string dtype by @galipremsagar in #20890
• Change CUDF_STRING_DTYPE to pandas.StringDtype for pandas 3.0 by @mroeschke in #21062
• Adjust cuDF grouby unit tests for existing pandas object types in pandas 3 by @mroeschke in #21081
• Fix StringDtype dask serialization in pandas 3 by @mroeschke in #21103
• Avoid object dtype in ColumnAccessor.to_pandas_index for pandas 3 by @mroeschke in #21110
• Move CategoricalDtype higher in cudf.dtype elif chain for pandas 3 by @mroeschke in #21129
• Match groupby idxmin/max skipna=True pandas 3.0 behavior with all null groups by @mroeschke in #21189
• Map numpy dtypes to new default string dtype in np_dtypes_to_pandas_dtypes for pandas 3 by @mroeschke in #21201
• Fix test_string_char_types for pandas 3 by @mroeschke in #21210
• Adjust more misc tests for pandas 3 by @mroeschke in #21228
• Preserve StringDtype correctly in StringColumn.to_pandas for pandas 3 by @mroeschke in #21206
• Update pandas pinnings in conda-forge recipes for pandas 3 by @mroeschke in #21302
• Update cudf_pandas proxies for pandas 3 by @mroeschke in #21606
• Change date_range time unit inferencing logic by @galipremsagar in #21694
• Make copy-on-write the default for pandas3 by @vyasr in #21801
• Fix multiple repr issues in cudf with pandas-3.0 by @galipremsagar in #21778
• Fix replace API & pytests by @galipremsagar in #21938
• Fix cudf.pandas pytest failures by @galipremsagar in #22137
• Fix setitem dtype validation, NaT column inference, and boolean-dtype join by @galipremsagar in #22255
• Reject pd.NA string-to-object casts in pandas-compatible mode by @galipremsagar in #22295
• Align groupby apply/rolling defaults with pandas 3 by @galipremsagar in #22448
• Align cudf find_common_type with pandas for empty/bool/datetime+timedelta inputs by @galipremsagar in #22466
• Add validations in various code-paths by @galipremsagar in #22336
• Merge pandas3 feature branch into main by @galipremsagar in #22518
• Forward-merge release/26.06 into main by @madsbk in #22585
• Remove deprecated .values_host property in cudf by @mroeschke in #22649
• Remove deprecated nvtext::edit_distance_matrix by @davidwendt in #22644
• Fix capture group handling in cudf::strings::findall by @davidwendt in #22408
• Removed deprecated multi-pattern overload of replace_re by @davidwendt in #22639
• Skip sum updates once SUM_WITH_OVERFLOW flag is set by @PointKernel in #22696
• Remove deprecated strings::like string_scalar API by @davidwendt in #22678
• Fix Categorical astype for nulls in pandas-compatible mode by @galipremsagar in #22652
• [cudf_streaming] Using rapidsmpf spill unspill partition by @nirandaperera in #22810
• Change behavior for regex class pattern [a-m-z] by @davidwendt in #22601
• Fix groupby any/all on null-containing string columns by @galipremsagar in #22926
• Align cudf indexing/setitem validation and alignment with pandas by @galipremsagar in #22912
• refactor(streaming): flatten namespaces and rename to snake_case by @vyasr in #22909
• [cudf_streaming] rapidsmpf backrefmixin API changes by @nirandaperera in #22911
• Refactor packed metadata to use an explicit table header by @madsbk in #22951
• Support skipna in groupby reductions (first/last, sum/prod/mean/median/min/max, idxmin/idxmax) by @galipremsagar in #22925
• Make pylibcudf.Table.columns() return a tuple by @madsbk in #23040
• Align DataFrame.reindex dtype and validation behavior with pandas by @galipremsagar in #22914
• Match pandas suffix handling in merge (None/non-string suffixes, invalid-container TypeError) by @galipremsagar in #23059
• Rapidsmpf logger API changes by @nirandaperera in #22972
• Refactor OrderScheme to track multiple Ordering descriptions by @rjzamora in #22896
• Return null for non-participating regex capture groups in strings::extract by @galipremsagar in #23123
• wheels: build CUDA 13 wheels with latest CTK (13.3.0) by @jameslamb in #23174
• Support explicit row counts for zero-column tables by @madsbk in #22765
• Fix MultiIndex indexing/.loc to match pandas by @galipremsagar in #22966
• Match pandas merge/join dtype, index and error semantics by @galipremsagar in #23060
• Rapidsmpf backref API changes for Host and Pinned MR by @nirandaperera in #23128
• Rewrite mixed inner/left/full join with post-filtering by @PointKernel in #23012
• Fix DataFrame.stack level resolution, ordering, and dtype preservation by @galipremsagar in #23370
• Rework approach to cudf-streaming bloom filter sizing by @wence- in #23067
• Match pandas ordering, dtypes, and validation in unstack/_pivot by @galipremsagar in #23368
• JNI bridge for the retention filter support for deletion vectors by @jihoonson in #23423

🐛 Bug Fixes

• Add back Index.fillna by @mroeschke in #21002
• Account of pandas.NamedAgg subclass change in pandas 3 by @mroeschke in #21023
• Account for non-fixed pandas.offsets.Day in date_range for pandas 3 by @mroeschke in #21036
• Fix columns validation logic of DataFrame.from_records by @mroeschke in #21034
• Unify _is_null_host_scalar into is_na_like by @galipremsagar in #21079
• Fix datetime binops issues in pandas-3.0 by @galipremsagar in #21172
• Fix convert_dtypes with new StringDtype in pandas 3 by @mroeschke in #21200
• Preserve StringDtype variants from pandas inputs for pandas 3 by @mroeschke in #21208
• Re-enable dtype=category in pandas-3.0 branch by @galipremsagar in #21604
• Fix all string type bugs by @galipremsagar in #21609
• Fix astype `Categori...

v26.06.01

What's Changed

🛠️ Improvements

• Temporarily WAR rapidsmpf<->libcudf conda package build cycle (#22520) by @jolorunyomi in #22772

Full Changelog: v26.06.00...v26.06.01

v26.06.00

What's Changed

🚨 Breaking Changes

• Undeprecate the byte-pair-encoding APIs by @davidwendt in #21760
• [Multi-GPU Polars] Introduce Ray mode for multi-GPU cudf-polars execution by @madsbk in #21746
• Get rid of relaxed constexpr across libcudf by @PointKernel in #21703
• [Multi-GPU Polars] Use current rmm resource in SPMD mode by @madsbk in #21842
• Remove obsolete statistics infrastructure by @rjzamora in #21857
• [Multi-GPU Polars] Create engines directly instead of factory functions by @madsbk in #21898
• Handle integers in floor division and power AST operators by @mhaseeb123 in #21831
• Enforce cudf_polars cardinality_factor and scheduler deprecations by @mroeschke in #21988
• [Multi-GPU Polars] Unify streaming engine options by @madsbk in #21930
• [Multi-GPU Polars] Split PDSH utils into legacy and new frontend paths by @madsbk in #21941
• Remove CUDAStreamPolicy enum and simplify CUDA stream policy by @vyasr in #22086
• [Multi-GPU Polars] Bind workers to topology-local hardware by @madsbk in #22113
• [FEA] Support Multi-Output JIT Transforms by @lamarrr in #21704
• Migrate RMM usage to CCCL MR design by @bdice in #22008
• Refactor cudf-polars plugin for Polars' test suite by @madsbk in #22301
• Remove legacy Dask-based streaming backends by @madsbk in #22358
• Make RapidsMPF the default runtime for cudf_polars streaming executor by @mroeschke in #22281
• Bump minimum Polars version to 1.35 by @mroeschke in #22459
• Introduce a process-wide singleton engine for .collect(engine="gpu") by @madsbk in #22410
• Remove cudf-polars[rapidsmpf] pip extra & numpy as a [test] dependency; add [dask] pip extra by @mroeschke in #22480
• Untangle target_partition_size and broadcast_join_limit by @rjzamora in #22411
• Replace --executor with extended --frontend choices in cudf-polars benchmarks by @madsbk in #22504
• Clean up legacy test scaffolding in cudf-polars by @madsbk in #22535
• [cudf_polars] Reorganize package layout by @madsbk in #22491
• Move collectives module by @rjzamora in #22578

🐛 Bug Fixes

• Fix TypeError when gathering on empty indices by @jberg5 in #21705
• CPU-only importable pdsh benchmark file by @TomAugspurger in #21791
• Change more Rapidsmpf Shuffler.wait_on to Shuffler.wait by @mroeschke in #21798
• Add missing headers to reader_impl_chunking_utils.cu by @bdice in #21784
• Fix TPC-DS query validation failures due to nulls_last mismatch by @Matt711 in #21814
• IWYU to fix latest CCCL compilation by @vyasr in #21839
• Fix additional dictionary tests to handle unordered keys by @davidwendt in #21773
• Add missing includes for cuda::std::abs by @PointKernel in #21845
• Fix segment calculation in TPC-DS Q54 by @Matt711 in #21829
• Workaround sum(nulls) difference between DuckDB and Polars in TPC-DS Q64 by @Matt711 in #21826
• Fix ambiguous stream constructor by @bdice in #21881
• Ensure cudf.pandas proxy object tests populate test-local type maps by @mroeschke in #21879
• Dont allow rtxpro6000 runners to pick up CI jobs by @Matt711 in #21954
• Fix union actor deadlock when input branches share a fanout by @Matt711 in #21949
• Fix expression decomposition when mixing fusable and non-fusable reductions by @Matt711 in #21822
• Fix type mismatch in groupby-count with multiple partitions by @Matt711 in #21934
• Return null instead of nan for pl.Expr.mean with rapidsmpf by @mroeschke in #21805
• Fix stream-ordering bugs related to pool streams by @vuule in #21908
• Fix data_alloc_size query bugs by @rjzamora in #21955
• Skip pinned memory tests on unsupported systems by @rjzamora in #21976
• Fix null_count incorrectly marked as pointwise by @vyasr in #21995
• Add sort_keys to benchmark validation for complex sort expressions by @Matt711 in #21817
• Avoid invalid pwise join when dynamic-planning is enabled by @rjzamora in #21977
• Validate PDS-DS Queries Q24, 47, 49, 94 by @Matt711 in #22007
• Fix validation failures in TPC-DS Q70 and Q79 by @Matt711 in #21820
• Fix OOM in PDS-DS Q78 by @Matt711 in #22009
• Workaround unsupported unary function in a groupby context in PDS-DS Q94 by @Matt711 in #22013
• Fix unreachable else branch in gather bitmask logic by @eternallyproud in #21946
• Fix cuda error when sorting empty pl.concat result by @jberg5 in #21825
• Cast groupby sum of integers result to schema in cudf_polars by @mroeschke in #21990
• Fix RTX PRO 6000 Blackwell CI by @bdice in #21999
• Exclude value_counts as a pointwise UnaryFunction in cudf_polars by @mroeschke in #22001
• Fix ast return_type_functor to handle decimal types with non-zero scale by @davidwendt in #21996
• Fix remote IO in cudf-polars pdsh benchmark by @ncclementi in #22090
• Fix deprecation warnings for set_as_build_table by @davidwendt in #22087
• Fix libcudf gather segfault in set_all_valid_null_masks by @davidwendt in #22092
• Fix stable ID for DataFrameScan by @rjzamora in #22091
• Fix PDS-H decimal validation failures by @Matt711 in #22107
• Optimize PDS-DS Q74 by @Matt711 in #22109
• Pass BufferResource for stream lifetime in rapidsmpf integration layer by @vyasr in #22110
• Pass required br argument to TableChunk.from_pylibcudf_table by @pentschev in #22116
• Fix missing rapidsmpf hiding real ImportError in benchmark scripts by @pentschev in #22114
• Fix partitioning metadata preservation for GroupBy by @rjzamora in #22111
• Expand CSE placeholders during HStack lowering by @rjzamora in #21796
• Fix lists::segmented_gather to return empty for empty input by @davidwendt in #22115
• Ensure insert_finished() is called on error paths for streaming collectives by @Matt711 in #22142
• Skip flaky upstream polars deadlock test by @Matt711 in #22182
• Fix CSE HStack lowering to respect with_columns semantics by @Matt711 in #22184
• CUDA 13.2 support: prefer __syncthreads() to block.sync() for shared memory fencing, fix compiler errors in C++ tests by @jameslamb in #22152
• Prevent memory corruption in ORC reader by @vuule in #22186
• Pin to pyarrow<24 in type checking environment by @TomAugspurger in #22230
• Pin PyArrow to <24 by @KyleFromNVIDIA in #22236
• Increase tolerance in test_groupby_categorical_key by @pentschev in #22249
• Set memory limit for DaskEngine by @TomAugspurger in #22242
• Revert Date casts in pdsh benchmarks by @TomAugspurger in #22232
• Fix flaky tracing test in cudf-polars by @TomAugspurger in #22012
• Pin polars version in type-checking environment by @TomAugspurger in #22256
• Fix nvbench handling of memory-resource objects by @davidwendt in #22257
• Temporarily increase max days without success to 40 days by @pentschev in #22264
• fix(cmake): exclude zstd, roaring, and cuco from install by @vyasr in #22263
• Rescale timestamp stats to the target precision in parquet pred...

v26.04.00

What's Changed

🚨 Breaking Changes

• Make offset vectors consistent across partitioning APIs by @PointKernel in #20937
• Remove unneeded parameters from make_dictionary_column by @davidwendt in #21032
• Deprecate nvtext::byte_pair_encoding APIs by @davidwendt in #21157
• Require pyarrow 19 or greater by @vyasr in #21183
• Accept device spans instead of buffers in hybrid scan APIs. by @mhaseeb123 in #21026
• Deprecate nvtext::edit_distance_matrix API by @davidwendt in #21159
• Enable parquet readers to select top-level cols by index by @mhaseeb123 in #21113
• Hybrid scan APIs to accept an mr to allocate output tables by @mhaseeb123 in #21263
• Add pylibcudf interface to cudf::round_decimal by @davidwendt in #21332
• Use the new make_table_chunks_available_or_wait API from RapidsMPF by @madsbk in #21291
• [BUGFIX] Make cudf's context mt-safe on initialization by @lamarrr in #21008
• Enforce thread safety in sort-merge join APIs by @shrshi in #21314
• Change dictionary columns to support unordered keys by @davidwendt in #21058
• Add cudf::column_nans_to_nulls API by @davidwendt in #21331
• Relocate distinct_count and unique_count from stream_compaction to reduction by @PointKernel in #21372
• Change missing value repr for string from to None by @mroeschke in #21425
• Make device_span an alias of cuda::std::span by @PointKernel in #21283
• Preserve pandas nullable types in dtype= arguments by @mroeschke in #21499
• [BUGFIX] Fix handling of scalar inputs in transform and row_ir by @lamarrr in #21443
• Preserve pandas nullable types from pandas objects by @mroeschke in #21514
• Remove unneeded stream/mr parameters from make_lists_column by @davidwendt in #21548
• Bump minimum NumPy dependency to 1.26 by @mroeschke in #21607
• Prune uneeded dependencies from pyproject.toml's and recipe.yml's by @mroeschke in #21608
• Adapt to rapidsmpf changes splitting communicator and streaming context by @wence- in #21632
• RapidsMPF SPMD mode by @madsbk in #21619
• Improve approx_distinct_count to work with preallocated sketch by @PointKernel in #21396
• [BUG] [cudf-polars] Remove rapidsmpf concat insert option by @nirandaperera in #21761
• Bump librdkafka to >=2.11.1,<2.12.0 by @mroeschke in #21951

🐛 Bug Fixes

• Add missing synchronisation in decode_delta_byte_array_kernel by @wence- in #21051
• Fixes index-based filtering of left join indices by @shrshi in #20990
• Fix issues in CSV reader benchmarks by @vuule in #21025
• Fix .first/last cudf_polars expr with empty columns by @mroeschke in #21089
• Switch date type in pandas PDSH queries by @Matt711 in #21101
• Preserve partitioning information in Cache nodes by @rjzamora in #21098
• Use Dask "context" to populate rapidsmpf context for proper statistics by @quasiben in #21126
• Change output dtype to int32 for datetime component methods by @Matt711 in #21099
• Update column names in pdsh query 18 by @TomAugspurger in #21119
• Preserve NA values in decimal ArrowExtensionArray comparisons by @Matt711 in #21168
• Handle empty format string in dt.strftime for GPU engine in cudf_polars by @mroeschke in #21143
• Handle Greek final sigma in str.lower() by @Matt711 in #21154
• Preserve type metadata in unary ops by @Matt711 in #21170
• Fix the host-device access to device_span data in cumulative_centroid_weight_grouped by @vuule in #21163
• Make JNI partition output consistent with cuDF by @zpuller in #21029
• Ensure empty windows return NaN for rolling aggregations by @Matt711 in #20677
• Fix at/iat indexers to reject list-like keys by @Matt711 in #21213
• Vendor Pandas' to_xarray in cudf.pandas by @Matt711 in #21175
• Raise error for string slicing on non-monotonic DatetimeIndex by @Matt711 in #21218
• Add __array_priority__ to proxy scalar types to match pandas by @Matt711 in #21235
• Change find_first_set_bit_kernel to use BlockReduce instead of cg::reduce by @davidwendt in #21234
• Add noinline declaration to secondary type-dispatching row-operators in Debug build by @davidwendt in #21197
• Fix casting to StringDtype for ArrowExtensionArray duration and decimal types by @Matt711 in #21230
• Fix pl.Expr.concat_str with boolean columns in cudf_polars by @mroeschke in #21145
• Only serialize column slice by @pentschev in #21301
• Fix return type of str.ip2int to uint32 by @mroeschke in #21324
• Fix integer overflow in cudf::hash_partition by @wence- in #21298
• Remove pandas compatibility gating for dtypes in datetime and timedelta columns by @galipremsagar in #21252
• Fix how cudf-polars tracing finds dataframe inputs by @TomAugspurger in #20647
• Use ColumnBase.create in from_arrow, follow-ups in pyarrow 19 decimal support by @mroeschke in #21317
• Fix validation for datetimes coming from arrow by @vyasr in #21369
• Fix n_workers detection in tpc-h benchmark by @TomAugspurger in #21368
• Fix list literal translation by @Matt711 in #20801
• Fix exponentially weighted mean with leading NULLs by @Matt711 in #20556
• Fix various issues in Python microbenchmarks by @vyasr in #21340
• Allow np.dtype(str) to coerce to object type for cudf column construction by @mroeschke in #21403
• Remove 4-byte padding requirement from the Brotli decompressor by @vuule in #21392
• Replace thrust::get with cuda::std::get by @mhaseeb123 in #21416
• Tests all APIs docstrings in doctests by @galipremsagar in #21404
• [BUG] Fixed a bug in the 10min notebook by @drivanov in #21400
• Fix empty string generation in get_json_object by @vyasr in #21418
• Save duckdb pdsh results to output file by @Matt711 in #21391
• Pin cupy<14 for wheel-tests-cudf and cudf-pandas unit tests by @Matt711 in #21469
• Fix generic_scalar_device_view::value() to honor decimal scale by @davidwendt in #21447
• Add notes to failing polars test by @vyasr in #21488
• Support CuPy 14 by @Matt711 in #21504
• Fix misaligned memory access in groupby by @PointKernel in #21513
• Deterministic build and probe table generation for join benchmarks by @shrshi in #21410
• Fix pds-h benchmark with --suffix by @TomAugspurger in #21520
• Fix OOB read in chunked Parquet reads with skip_rows/num_rows by @vuule in #21407
• rapidsmpf compatibility in unspill_partitions by @TomAugspurger in #21539
• Handle edge case dectection in pds-h validation by @TomAugspurger in #21544
• Fix scan/ewm logic to pass stream to column_device_view::create by @davidwendt in #21557
• Account for the rank index in streaming Sink by @rjzamora in #21555
• Fix decimal precision loss in cast: float->decimal by @Matt711 in #21450
• Fix JIT filtering of join indices to properly handle nulls in nullable columns by @shrshi in #21570
• Fix span index-out-of-range error/assert in parquet::decode_page_data_generic kernel by @davidwendt in #2157...

v26.02.01

What's Changed

🐛 Bug Fixes

• Backport #21301: Only serialize column slice by @pentschev in #21328

Full Changelog: v26.02.00...v26.02.01

v26.02.00

What's Changed

🚨 Breaking Changes

• Avoid counting nulls and creating null mask in groupby aggregation MERGE_M2 by @ttnghia in #20716
• Remove cudf::get_current_device_resource by @bdice in #20688
• Avoid creating null mask in groupby aggregation M2 by @ttnghia in #20726
• Remove deprecated left semi- and anti- join APIs by @shrshi in #20668
• Inline and simplify some column methods by @vyasr in #20819
• Enable copy-on-write in cudf.pandas by @vyasr in #20401
• [FEA] Improve Null-Aware Operator Support in AST-Codegen by @lamarrr in #20206
• Remove legacy hash-combine logic and unify hashing with row hasher by @PointKernel in #20796
• Remove deprecated .from_pandas constructors by @mroeschke in #20925
• Remove deprecated Series.data by @mroeschke in #20914
• Remove all base attributes from ColumnBase by @vyasr in #20961
• Fix handling of unquoted strings in the CSV reader by @vuule in #20996

🐛 Bug Fixes

• Avoid duplicate streaming nodes for the rapidsmpf runtime by @rjzamora in #20586
• Handle scalar arguments in ternary expression by @Matt711 in #20600
• fix(noarch): use noarch build script in noarch build by @gforsyth in #20654
• fix(conda): matrix out noarch builds by cuda-major version by @gforsyth in #20678
• Include RMM in type checking environment and update type annotations for optional stream by @TomAugspurger in #20636
• Add no-op path for ArrowExtensionArray.astype by @Matt711 in #20580
• Skip pytorch integration tests if CUDA is not available by @Matt711 in #20729
• Always delay CUDA Array Interface pointer access by @vyasr in #20719
• Fix various copy-on-write bugs by @vyasr in #20744
• Fix leaks in cuDF java tests by @abellina in #20767
• Fix plc.Scalar.from_py(datetime.datetime) incorrectly localizing naive datetimes by @mroeschke in #20769
• Don't remove double casts in cudf_polars by @mroeschke in #20773
• Fixes struct column handling in sort-merge joins by @shrshi in #20664
• Fix for synccheck compute-sanitizer errors across Parquet gtest by @mhaseeb123 in #20775
• Pin numpy<2.4.0a0 in mypy pre-commit environment by @TomAugspurger in #20781
• Raise when trying to run queries on different devices in same process by @wence- in #20617
• Ensure min_periods=0 is passed through rolling aggregations by @Matt711 in #20653
• Fix racecheck errors in the ORC reader by @vuule in #20792
• Fix the crash of multi-threaded parquet reader benchmark by @kingcrimsontianyu in #20783
• Fix racecheck reported by DATA_CHUNK_SOURCE_TEST in inflate_kernel by @davidwendt in #20804
• Fix racecheck in the gpu_debrotli_kernel by @davidwendt in #20806
• Ensure literal groupby aggregations are broadcasted to key length in cudf_polars by @mroeschke in #20776
• Pin aiobotocore<3 to fix CI failures by @TomAugspurger in #20844
• Fix racecheck in parquet decode_page_data_generic kernel by @davidwendt in #20850
• Avoid generating empty TableChunks in streaming scan nodes by @rjzamora in #20815
• Fix dask imports in CudfFusedParquetIOHost by @rjzamora in #20845
• Fix UB due to OOM Exception in ParquetReaderTest.ManyLargeLists by @lamarrr in #20841
• Fix racecheck/synccheck in JSON parse_fn_string_parallel kernel by @davidwendt in #20856
• Fix racecheck in ORC decode_column_data_kernel by @davidwendt in #20853
• Disable flatbuffers tests in CMake configuration by @bdice in #20848
• Upper bound on aiosqlite in polars-upstream job by @TomAugspurger in #20866
• Fix boolean casting consistency with Pandas (#20746) by @aryansri05 in #20747
• Add retries to requests made to PyPI's JSON API by @TomAugspurger in #20865
• Fix size_type overflow in multiple APIs by @vuule in #20857
• Fix racecheck in parquet compute_string_page_bounds_kernel by @davidwendt in #20868
• Fix dictionary::encode to honor indices-type parameter by @davidwendt in #20842
• Add missing headers to row_ir.hpp, row_ir.cpp by @bdice in #20834
• Fix parquet_options in pdsh benchmark by @TomAugspurger in #20893
• Add stream synchronize to tdigest generate_group_cluster_info by @davidwendt in #20846
• Only install RMM in mypy env on linux by @TomAugspurger in #20878
• Make nvcomp export unconditional by @vyasr in #20828
• Ensure we have nvjitlink from the CUDA version used at build time or newer and upgrade numba-cuda lower bound by @bdice in #20873
• Fix size_type overflow in the ORC writer by @vuule in #20889
• Constrain pyparsing version by @vyasr in #20935
• Revert #20902 by @vyasr in #20955
• Add force-blocking-launches to run_compute_sanitizer_test script by @davidwendt in #20962
• Fix racecheck error in parquet delta_byte_array_decoder::string_scan by @davidwendt in #20967
• Fix racechecks reported in parquet gpuEncodePages kernel by @davidwendt in #20975
• Don't encode s3 paths for kvikio_remote_io in read_json by @mroeschke in #20976
• Allow sort merge join to go above int32 output row limits by @revans2 in #20960
• Correct stream ordered deallocation in Join by @TomAugspurger in #20981
• Reintroduce Buffer.nbytes property by @pentschev in #21027
• Fix SHA hash OOB on strings that are exact multiples of message chunk size by @rishic3 in #21004
• Temporarily disable IWYU for nightly tests by @davidwendt in #21045
• Fix cudf-polars multi-partition distributed sort by @TomAugspurger in #21047
• Backport #21051 by @wence- in #21086
• Pin pandas for pylibcudf testing by @galipremsagar in #21124
• Hide pinned pool instantiation to avoid symbol conflicts with nvcomp by @vyasr in #21161
• Specialize field type checking for bool in Parquet thrift list decoder by @mhaseeb123 in #21144
• Fix reading of CSV files with double quotes in unquoted strings by @vuule in #21151
• Revert the multithreaded optimization in the CSV reader by @vuule in #21198
• Pin sqlglot in third-party integration tests by @Matt711 in #21271
• Exclude sqlglot version 28.7 from CI by @Matt711 in #21293

📖 Documentation

• Add note to developer guide about null values being undefined by @bdice in #20645
• [DOC] Add cudf-polars to the example build command by @Matt711 in #20763
• Clarify internal API header placement guidelines for details headers by @PointKernel in #20985
• Clarify deprecation message for cudf::round by @nirandaperera in #20809
• Require nvcc 12.9 in contributing guide by @bdice in #21186

🚀 New Features

• Expose cudf::compute_column_jit to python by @Matt711 in #20697
• Add configuration option for max-io-threads by @quasiben in #20606
• Return stats from lower_ir_graph by @rjzamora in #20528
• Promote join_kind from detail namespace to public by @PointKernel in #20703
• Make DataFrameScan and DataFrameSourceInfo pickle-able by @rjzamora in #20732
• Add compute-sanitizer dispatch action by @bdice in #20542
• Add RapidsMPF Al...

v25.12.00

What's Changed

🚨 Breaking Changes

• Rewrite JNI functions to use JNI_TRY/JNI_CATCH by @ttnghia in #19053
• Remove compatibility with nvCOMP versions before 5.0 by @vuule in #20140
• Remove DataFrame.apply_chunks, Groupby.apply_grouped by @mroeschke in #20194
• Change .str.starts/endswith with tuple argument to match any pattern instead of pairwise matching by @mroeschke in #20249
• [cudf-polars] CUDA stream by @madsbk in #20154
• Chunked read parquet, prepend index column, and apply deletion vector by @mhaseeb123 in #20201
• Zero-copy hostdevice_vector on integrated systems by @vuule in #20225
• Use int64_t for the num_rows slot in parquet_reader_options by @wence- in #20256
• Require CUDA 12.2+ by @jakirkham in #20416
• Remove compatibility for CCCL < 3.1 by @bdice in #20468
• Remove deprecated types and APIs by @vuule in #20422
• Support signed integers and decimals in SUM_WITH_OVERFLOW groupby by @PointKernel in #19598
• Change groupby-scan COUNT to 1-based results by @davidwendt in #20168
• Change strings::like() pattern parameter from string_scalar to string_view by @davidwendt in #20428
• No-op performance tracking wrappers by @galipremsagar in #20595

🐛 Bug Fixes

• Copy attrs at correct place in DataFrame constructor by @galipremsagar in #20074
• Handle missing nightly runs in pandas tests job by @galipremsagar in #20081
• Fix numpy ufunc for DataFrame by @galipremsagar in #20070
• Unproxy few unnecessary testing utilities in pandas by @galipremsagar in #20088
• Fix libcudf groupby benchmarks to not include internal cache by @davidwendt in #20038
• Fix cudf.date_range with non-iso start and end date strings by @mroeschke in #20116
• Fix create_distinct_rows_column to create non-nullable columns by @davidwendt in #20082
• Fix arrow timestamp frequency cases in cudf.pandas by @galipremsagar in #20128
• Cast inputs to true division from decimal to float by @Matt711 in #20077
• Handle NVMLError_NotSupported in cudf-polars by @TomAugspurger in #20179
• Fix RMM JNI pinned_fallback_host_memory_resource for CCCL 3.1.0 by @bdice in #20160
• Require passing memory resources to from_libcudf methods by @vyasr in #20171
• Enable hash-groupby for decimal32/64 type and MEAN aggregation by @davidwendt in #20040
• Align decimal dtypes in predicate before conditional join by @Matt711 in #20060
• Change stream_checking_resource_adaptor::do_deallocate to noexcept by @vyasr in #20218
• Deallocation should be noexcept by @bdice in #20219
• Fix a race condition in the decode of delta encoded Parquet columns by @vuule in #20216
• Fix the host-device tdigest offsets by using cuda::std::span by @PointKernel in #20220
• Add stream and mr arguments to Column.from_arrow type stub by @TomAugspurger in #20244
• Pin deltalake in cudf-polars-polars-tests CI job by @TomAugspurger in #20255
• Pin ibis-framework<11.0.0 by @Matt711 in #20267
• Add private attributes for cudf.pandas proxy objects by @galipremsagar in #20276
• Add Proxy for SparseAccessor by @galipremsagar in #20278
• We need this to pacify mypy by @wence- in #20285
• Purge non-empty nulls for the generated lists columns in data generation utility by @ttnghia in #20283
• Fix missing table compatibility check in two_table_comparator constructor by @PointKernel in #20305
• Fix the check for equal num_cols across empty parquet sources by @mhaseeb123 in #20320
• Add nans_to_nulls to Frame by @galipremsagar in #20314
• Add support for list type in get by @galipremsagar in #20332
• Fix decimal dtype serialization in cudf-polars by @Matt711 in #20300
• Make the GroupedRollingWindowexpression node reconstructable in cudf-polars by @Matt711 in #20288
• Ensure pylibcudf.Scalar.from_py uses CUDA streams by @TomAugspurger in #20340
• Skip failing cudf-polars test due to hash groupby bug by @Matt711 in #20356
• Support order by keys for order-sensitive scalar aggregations in grouped windows by @Matt711 in #20350
• Honor user-passed stream in slice_strings for scalar inputs by @mroeschke in #20349
• Thread missing streams in column/table view creation to char size calculation by @vyasr in #20351
• Fix missed-sync for mapping_indices_kernel in hash-based groupby aggregation by @ttnghia in #20370
• Fix a few SPDX-related issues by @KyleFromNVIDIA in #20364
• Fix a dtype bug in column constructor by @galipremsagar in #20384
• Refactor as_column dtype parameter calls by @galipremsagar in #20379
• Add CUDA stream to cudf_polars.Column.deserialize by @TomAugspurger in #20396
• Add missing CUDA stream to cudf-polars left-semi join by @TomAugspurger in #20398
• Fix various string APIs to work with extension types by @galipremsagar in #20368
• Add parameter validation for merge and MultiIndex.from_frame by @galipremsagar in #20382
• Fix nvtext::normalize_characters special token case by @davidwendt in #20242
• Fix pinned memory resource shared_pointer lifetime in tests. by @bdice in #20407
• Support new nvcompStatus_t enum value by @vuule in #20376
• Don't skip blank CSV lines rows after the header in cudf-polars scan_csv by @mroeschke in #20341
• Fix OOB accesses in JSON_CornerCase_Empty test and get_row_array_parent_col_id function by @bdice in #20421
• Change calls to cudaMemcpyToSymbol to cudaMemcpyToSymbolAsync by @davidwendt in #20374
• Do not accelerate pandas._config.config by @Matt711 in #20413
• Return timedelta instead of datetime type with std with datetime type with missing values by @mroeschke in #20439
• Disallow non-bool skipna arguments to reduction methods by @mroeschke in #20436
• Fix parquet scans for duckDB PDS-DS by @Matt711 in #20388
• Support __array_function__ on the proxy array type by @Matt711 in #20419
• Make memory_usage and __sizeof__ proxy attributes and always skip all memory usage tests by @Matt711 in #20425
• Add input validation for from_records by @galipremsagar in #20412
• Use computed reduction result type for empty sum and product aggregations by @mroeschke in #20438
• Correct level arg validation for Index.isin, unique by @mroeschke in #20449
• Add private _grouper attribute to DataFrameGroupBy proxy type by @Matt711 in #20448
• Raise ValueError when indexing with zero step slice by @mroeschke in #20453
• Raise IndexError for float-like indexers in RangeIndex/MultiIndex.getitem by @mroeschke in #20454
• Disallow slice(bool, ...) in DataFrame.loc with MultiIndex by @mroeschke in #20457
• Fix core dump in MemoryCleaner by @res-life in #19872
• Disallow multiple ellipse values in loc/iloc indexing by @mroeschke in #20456
• Fix scan operations for string columns by @galipremsagar in #20460
• Fix UTF8 data generator in libcudf benchmarks utility by @davidwendt in #20465
• Handle dealloc in stream-ordered cudf-polars ops by @TomAugspurger in #20467
• Raise on unsupported unsta...

v25.10.00

🚨 Breaking Changes

• Remove UCX-Py (#19979) @pentschev
• Revert "Migrate mixed join to use multiset #19660" (#19933) @PointKernel
• Fill missing values in Series/Index.values for numeric types with np.nan by default (#19923) @mroeschke
• Remove deprecated DataFrame.apply_rows, deprecate DataFrame.apply_chunks and Groupby.apply_grouped (#19896) @mroeschke
• Move prefetching out of experimental and simplify the API (#19875) @vyasr
• Add join *_match_context APIs to hash join (#19835) @PointKernel
• Vendor libnvcomp in libcudf (#19743) @bdice
• Migrate mixed join to use multiset (#19660) @PointKernel
• Separate row mask and page mask computation and usage (#19537) @mhaseeb123
• [FEA] Implement null-aware transforms and filters (#19502) @lamarrr
• Support output-type for MEDIAN/QUANTILE aggregation in cudf::reduce (#19267) @davidwendt

🐛 Bug Fixes

• Fix edge cases in statistics collection (#20094) @rjzamora
• Fix multi-partition Filter bug (#20075) @rjzamora
• Fix reindex to fill only the reindexed values with fill_value (#20063) @galipremsagar
• Fix arrow arrays + numpy ufunc interaction (#20047) @galipremsagar
• Fix race conditions in ORC reader decimal decoding (#20044) @vuule
• Keep mr alive along with arrow tables and columns (#20028) @vyasr
• Fix value_counts missing nan bug (#20026) @galipremsagar
• Compatibility for rapidsmpf's unspill_partitions (#20020) @TomAugspurger
• Fix type metadata preservation in shift (#20017) @galipremsagar
• Fix incorrect type propagation in dataframe assignment (#20010) @galipremsagar
• Fix OOB memory read in decode_page_data_generic kernel (#19995) @davidwendt
• Fix data_type creation in ast::operation::instantiate (#19994) @davidwendt
• Skip Narwhals pandas get_dtype_backend[pyarrow] tests after ArrowDtype proxy changes (#19992) @Matt711
• Make cudf.pandas callables usable with inspect.getfullargspec (#19988) @mroeschke
• Align decimal dtypes to schema after parquet IO scan (#19974) @Matt711
• Avoid undefined numpy protocols on cudf.pandas proxy objects (#19968) @mroeschke
• Skip failing polars iceberg test (#19955) @Matt711
• Revert "Migrate mixed join to use multiset #19660" (#19933) @PointKernel
• Define FrozenList proxy independently in cudf.pandas (#19931) @mroeschke
• Ignore scalars when broadcasting for horizontal string concatenation in cudf-polars (#19893) @Matt711
• Fix is_valid_rolling_aggregation for STD aggregation (#19888) @davidwendt
• Fix a decompression parameter in the chunked ORC reader (#19882) @vuule
• Skip flaky stats tests pending follow up (#19881) @brandon-b-miller
• Require list type for is_valid_aggregation and MERGE_LISTS/SETS (#19876) @davidwendt
• Temporary solution to ensure data-source/sink stream ordering (#19874) @kingcrimsontianyu
• Check for integer overflow in cudf::strings::find_multiple (#19867) @davidwendt
• Fix missing stream from cudf::top_k_order (#19866) @davidwendt
• Disallow loc.setitem with list-like indexer when list elements not in index (#19851) @mroeschke
• Fix .str.replace ignoring n for single character replacements (#19848) @mroeschke
• Fix strings::find_instance warp parallel logic (#19845) @davidwendt
• Add changed-files to the needs of every job that requires it (#19830) @Matt711
• xfail polars decimal(precision=None) test (#19821) @Matt711
• Fix empty column returned by cudf::from_arrow_stream_column (#19812) @davidwendt
• Filter pandas warning in dask_cudf test (#19808) @TomAugspurger
• Update identify_stream_usage CUDA runtime hooks to CUDA 13 (#19807) @robertmaynard
• When bundling libnvcomp.so.X only append the major version value (#19786) @robertmaynard
• Improvements to pylibcudf.from_iterable_of_py (#19781) @Matt711
• Avoid using multiple Cache nodes with the same hash (#19769) @rjzamora
• Fix window var() test failures from float rounding (#19761) @Matt711
• Use is_compressed field from Parquet V2 data page headers to determine if they are compressed (#19755) @mhaseeb123
• Fix bug in eval function with nvtx-0.2.11 (#19754) @galipremsagar
• Fix ndsh benchmarks nvtx range usage (#19753) @davidwendt
• Support nan in non-floating point column in cudf-polars (#19742) @Matt711
• Fix filter call in benchmark (#19732) @vyasr
• Suppress NVRTC warning from stdint.h (#19712) @davidwendt
• Correctly decode boolean lists in chunked parquet reader (#19707) @mhaseeb123
• Add new xfails for xarray release (#19705) @vyasr
• Fix "--executor" pytest parameter for cudf-polars (#19703) @rjzamora
• Match polars semantics for rolling-sum with all-null windows (non-empty) (#19680) @Matt711
• [BUG] Set query_set arg when validating/running cudf-polars PDS-DS benchmarks (#19674) @Matt711
• Fix group_by().agg() on non-aggregatable dtypes (#19669) @Matt711
• Fix broken links in 10min notebook (#19665) @Matt711
• Skip managed memory test if managed memory not supported in cudf-polars (#19653) @Matt711
• Fix integer overflow in warp-per-row grid calculation (#19638) @davidwendt
• Propagate exceptions thrown in async IO operations (#19628) @vuule
• Make DataFrame.dtypes not fallback to CPU always (#19627) @galipremsagar
• Set scalar to valid in range_window_bounds unbounded/current_row (#19622) @davidwendt
• Enable data page mask computation for nullable list and struct columns (#19617) @mhaseeb123
• Fix cudf::sequence() to throw exception for invalid scalar inputs (#19612) @davidwendt
• Fix uninitialized variable and misaligned write in parquet generic decoder (#19601) @mhaseeb123
• Compatibility with rapidsmpf 25.10.0 (#19591) @TomAugspurger
• Avoid querying device memory on systems without it in dask-cudf (#19577) @Matt711
• Avoid querying device memory on systems without it in cudf-polars benchmarks (#19575) @Matt711
• Increase alignment requirement for parquet bloom filter to 256 (#19573) @mhaseeb123
• Fix strftime with non-exact %a, %A, %b, %B (#19570) @mroeschke
• Fix OOB memcheck error in group_rank_to_percentage utility (#19567) @davidwendt
• Fix logic for number of unique values generated by data profile in benchmarks (#19540) @shrshi
• Fix contiguous-split nvbench cmake build (#19534) @davidwendt
• Fix value counts expression when the column has nulls (#19524) @Matt711
• Prefer Column.astype over plc.unary.cast in the fill null unary function expression (#19479) @Matt711
• Fix missing return in StringFunction.Strptime strict=True path (#19464) @Matt711
• Make dividing a boolean column return f64 dtype in cudf-polars (#19443) @Matt711
• branch-25.10-merge-branch-25.08 (#19429) @davidwendt
• Replace sprintf with std::format in libcudf parquet tests (#19364) @davidwendt

📖 Documentation

• Update missing docs (#19925) @vyasr
• Add examples of null handling to doxygen for cudf::rank (#19774) @davidwendt
• Fix cudf-polars dependency list docs (#19750) @pentschev
• Update cuDF classic testing documention regarding testing organization (#19745) @mroeschke
• Improve documentation around why we need no_gc_clear on pylibcudf Scalars (#19661) @vyasr

🚀 New Features

• Add memory resource parameters to interop, merge, and transpose (#20007) @vyasr
• Add mixed join benchmark with complex AST operators (#20004) @PointKernel
• Add memory resource arguments to join, round, and labeling (#20001) @vyasr
• cudf-polars strptime format inference (#19997) @brandon-b-miller
• Filter parquet row groups using byte offset bounds (#19991) @mhaseeb123
• Add memory resource arguments to concatenate (#19943) @vyasr
• Use column statistics to generate the physical plan in cuDF-Polars (#19940) @rjzamora
• Add all missing stream parameters (#19922) @vyasr
• Remote IO support in cudf-polars (#19921) @Matt711
• Add streams to io/timezone and io/text modules (#19913) @vyasr
• Add stream support to all nvtext modules (#19911) @vyasr
• Add streams to all top-level strings modules (#19910) @vyasr
• Update strings split APIs with stream parameters (#19909) @vyasr
• Support ordered grouped windows in cudf-polars (#19891) @Matt711
• Add local row-count and unique-count estimates to explain(... logical=True) (#19864) @rjzamora
• Add join *_match_context APIs to hash join (#19835) @PointKernel
• Support rank(...).over(...) expressions in cudf-polars (#19803) @Matt711
• Add strings to/from encoded integer APIs (#19789) @davidwendt
• Add to_arrow method to pylibcudf core types (#19787) @Matt711
• Add streams to strings convert APIs (#19780) @vyasr
• Add an option to support reading ORC timestamp column as UTC time. (#19773) @res-life
• Support null_count in groupby/rolling context (#19739) @Matt711
• Collect join-key information in cudf-polars (#19736) @rjzamora
• Add count aggregation support to cudf::reduce (#19734) @davidwendt
• [FEA] Implement AST Expression - JIT codegen (#19733) @lamarrr
• Add streams to all scalar factories (#19729) @vyasr
• Add streams to reshape (#19728) @vyasr
• Add streams to null mask APIs (#19727) @vyasr
• Add streams to column APIs (#19726) @vyasr
• Construct next-gen parquet reader with pre-populated footer (#19724) @mhaseeb123
• Require numba-cuda>=0.19.0,<0.20.0a0 (#19711) @brandon-b-miller
• Support over expression (window mapping) in cudf-polars (#19684) @Matt711
• Add streams support to all list APIs (#19683) @vyasr
• [FEA] Add Filter Benchmark (#19678) @lamarrr
• Add streams to pylibcudf join APIs (#19672) @vyasr
• Add streams to sorting APIs (#19671) @vyasr
• [FEA] Remove excessive copies of JITIFY's ProgramData during JIT kernel launch (#19667) @lamarrr
• Add streams to hashing APIs (#19663) @vyasr
• Use a more robust metric for sorting (de)compression tasks (#19656) @vuule
• Add streams support to datetime APIs (#19654) @vyasr
• Add streams to stream_compaction (#19651) @vyasr
• Enable casting pl.Datetime to integer types in cudf-polars (#19647) @brandon-b-miller
• Add Java JNI interface to get Gpu UUID (#19646) @res-life
• Add reduction with overflow detection (#19641) @PointKernel
• Upgrade to nvCOM...

[NIGHTLY] v25.12.00

🔗 Links

• Development Branch
• Compare with main branch

🚨 Breaking Changes

• Change .str.starts/endswith with tuple argument to match any pattern instead of pairwise matching (#20249) @mroeschke
• Remove DataFrame.apply_chunks, Groupby.apply_grouped (#20194) @mroeschke
• [cudf-polars] CUDA stream (#20154) @madsbk
• Remove compatibility with nvCOMP versions before 5.0 (#20140) @vuule
• Rewrite JNI functions to use JNI_TRY/JNI_CATCH (#19053) @ttnghia

🐛 Bug Fixes

• We need this to pacify mypy (#20285) @wence-
• Purge non-empty nulls for the generated lists columns in data generation utility (#20283) @ttnghia
• Add Proxy for SparseAccessor (#20278) @galipremsagar
• Add private attributes for cudf.pandas proxy objects (#20276) @galipremsagar
• Pin ibis-framework<11.0.0 (#20267) @Matt711
• Pin deltalake in cudf-polars-polars-tests CI job (#20255) @TomAugspurger
• Add stream and mr arguments to Column.from_arrow type stub (#20244) @TomAugspurger
• Fix the host-device tdigest offsets by using cuda::std::span (#20220) @PointKernel
• Deallocation should be noexcept (#20219) @bdice
• Change stream_checking_resource_adaptor::do_deallocate to noexcept (#20218) @vyasr
• Fix a race condition in the decode of delta encoded Parquet columns (#20216) @vuule
• Handle NVMLError_NotSupported in cudf-polars (#20179) @TomAugspurger
• Require passing memory resources to from_libcudf methods (#20171) @vyasr
• Fix RMM JNI pinned_fallback_host_memory_resource for CCCL 3.1.0 (#20160) @bdice
• Fix arrow timestamp frequency cases in cudf.pandas (#20128) @galipremsagar
• Fix cudf.date_range with non-iso start and end date strings (#20116) @mroeschke
• Unproxy few unnecessary testing utilities in pandas (#20088) @galipremsagar
• Fix create_distinct_rows_column to create non-nullable columns (#20082) @davidwendt
• Handle missing nightly runs in pandas tests job (#20081) @galipremsagar
• Cast inputs to true division from decimal to float (#20077) @Matt711
• Copy attrs at correct place in DataFrame constructor (#20074) @galipremsagar
• Fix numpy ufunc for DataFrame (#20070) @galipremsagar
• Align decimal dtypes in predicate before conditional join (#20060) @Matt711
• Enable hash-groupby for decimal32/64 type and MEAN aggregation (#20040) @davidwendt
• Fix libcudf groupby benchmarks to not include internal cache (#20038) @davidwendt

📖 Documentation

• Add profiling guide (#20292) @bdice
• Add note that --rmm-async only affects distributed scheduler. (#20129) @bdice

🚀 New Features

• Implement ARGMIN and ARGMAX aggregations for reduction (#20207) @ttnghia
• Add remaining memory resources (#20197) @vyasr
• Add memory resources to scalars (#20196) @vyasr
• Skip decompression of pruned parquet pages (#20192) @mhaseeb123
• Add memory resources to replace, json, and hashing (#20150) @vyasr
• Support decimal literals in cudf-polars (#20147) @Matt711
• Add pylibcudf is_valid_reduce_aggregation API (#20145) @davidwendt
• Add memory resources to I/O modules (#20136) @vyasr
• Add memory resources to reduce, column, column_factories, and contiguous_split (#20135) @vyasr
• Passthrough unary ops through Parquet predicate pushdown (#20127) @mhaseeb123
• Add memory resource to all strings modules (#20123) @vyasr
• Add memory resources to all nvtext APIs (#20119) @vyasr
• Add an example to inspect parquet files and dump row group and page level metadata information (#20117) @mhaseeb123
• Allow multiple calls to cudf::initialize and cudf::deinitialize (#20111) @vuule
• Remove rounding from cudf java (#20110) @pmattione-nvidia
• Add memory resources to groupby, datetime, and lists modules (#20102) @vyasr
• Add memory resources to search, reshape, and partitioning module (#20101) @vyasr
• Add memory resources to rolling, sorting, and quantiles modules (#20099) @vyasr
• Add memory resources to binaryop, copying, and stream_compaction (#20059) @vyasr
• Add memory resources to unary, transform, and filling modules (#20054) @vyasr
• Support cum_sum(...).over(...) expressions in cudf-polars (#19908) @Matt711
• Support forward/backward filling null values in a grouped window context (#19907) @Matt711
• [FEA] Implement JIT Filter for read_parquet (#19831) @lamarrr
• Add an example to demonstrate the use of next-gen parquet reader to read a parquet file with highly selective filters (#19469) @mhaseeb123
• Rewrite JNI functions to use JNI_TRY/JNI_CATCH (#19053) @ttnghia
• Add support for maintain_order param in joins (#17698) @Matt711

🛠️ Improvements

• Add more Python type annotations to cudf/core (#20287) @mroeschke
• Skip mypy in pre-commit.ci (#20286) @bdice
• Remove extraneous host_memory_resource include (#20284) @bdice
• Add numpy to the mypy pre-commit environment (#20282) @vyasr
• Add MultiIndex.dtypes (#20279) @galipremsagar
• Add more type annotations to cudf/core/column subclasses (#20277) @mroeschke
• Handle unordered grouped windows properly for null filling and cum sums (#20275) @Matt711
• Unpin DuckDB and Ibis in cudf.pandas thirdparty tests (#20269) @mroeschke
• Enable sccache-dist connection pool (#20264) @trxcllnt
• Add ability to set the source_info of parquet_reader_options (#20253) @wence-
• Update ConfigOptions for rapidsmpf-streaming integration (#20252) @rjzamora
• Add arm testing of cudf.pandas unit tests (#20251) @vyasr
• Add pylibcudf to pre-commit linting and fix outstanding errors (#20250) @vyasr
• Change .str.starts/endswith with tuple argument to match any pattern instead of pairwise matching (#20249) @mroeschke
• Move and rename ScanPartitionPlan (#20248) @rjzamora
• Standardize setting StructDtype field names post libcudf conversion (#20235) @mroeschke
• Prevent accidental copies of expensive-to-copy object types (#20226) @vuule
• More mypy and docs fixes (#20224) @vyasr
• Configuration for which metrics are enabled during tracing (#20223) @TomAugspurger
• Fix parquet row number check for page bounds (#20217) @pmattione-nvidia
• Rename comparison_binop_generator to arg_minmax_binop_generator and corresponding file to nested_types_extrema_utils.cuh (#20212) @Copilot
• Fix various typing errors (#20205) @vyasr
• Stop using libcudf default parameters in pylibcudf (#20204) @vyasr
• Pin pydantic<2.12 in ci/test_cudf_polars_polars_tests.sh (#20200) @mroeschke
• Support binops between float scalar to decimal column (#20199) @mroeschke
• Add an overhead field to cudf-polars tracing (#20198) @TomAugspurger
• Remove DataFrame.apply_chunks, Groupby.apply_grouped (#20194) @mroeschke
• [pre-commit.ci] pre-commit autoupdate (#20189) @pre-commit-ci[bot]
• Revert "Temporarily disable conda-java-tests" (#20184) @bdice
• Don't assume cudf_polars benchmarking scale factor is always an integer (#20182) @mroeschke
• Remove unnecessary work from read_parquet_metadata (#20180) @vuule
• Reduce execution times for parquet dictionary tests (#20176) @mhaseeb123
• Skip filtering Parquet row groups with dictionaries if there are non-dict encoded pages (#20175) @mhaseeb123
• Improve performance of groupby tdigests gtests (#20173) @davidwendt
• Update to rapids-logger 0.2 (#20172) @bdice
• Split row operator header (#20166) @PointKernel
• Add PDSH benchmark runner for cudf.pandas (#20164) @mroeschke
• Temporarily disable conda-java-tests (#20162) @bdice
• Manual forward merger for Branch 25.12 - branch 25.10 (#20157) @galipremsagar
• [cudf-polars] CUDA stream (#20154) @madsbk
• Avoid NumericalColumn call from CategoricalColumn.children (#20153) @mroeschke
• Branch 25.12 merge branch 25.10 (#20152) @vyasr
• Make ListColumn._transform_leaves convert via pylibcudf (#20151) @mroeschke
• Make ColumnBase.as_*_column convert via pylibcudf (#20149) @mroeschke
• Make ColumnBase.deserialize construct via pylibcudf (#20142) @mroeschke
• Remove unused ColumnBase.view (#20141) @mroeschke
• Remove compatibility with nvCOMP versions before 5.0 (#20140) @vuule
• Adjust rmm pool handling in PDSH benchmarks (#20138) @TomAugspurger
• Fix slowdown in cudf-polars distributed tests (#20137) @TomAugspurger
• Disable async MR priming in cudf.pandas (#20133) @bdice
• Fix type annotations in cudf-polars (#20131) @TomAugspurger
• Add tests for AUTO and HYBRID (de)compression modes (#20126) @vuule
• Run cudf-polars wheels unit tests with more than 1 process (#20124) @mroeschke
• Add pyarrow stubs to mypy environment and fix associated errors (#20118) @vyasr
• Avoid running pandas unit tests for private functionality with cudf.pandas (#20115) @mroeschke
• Remove MultiIndex.from_pandas pytest benchmark (#20112) @mroeschke
• Use 8 processes for pandas tests, show top 10 test times (#20109) @bdice
• Reduce verbosity of running the pandas test suite (#20107) @vyasr
• Switch host_vector and host_span dependency (#20106) @davidwendt
• Make Column.set_mask go through pylibcudf (#20103) @mroeschke
• Have ListColumn.from_sequence go through pylibcudf (#20098) @mroeschke
• Deprecate legacy public row operators (#20097) @PointKernel
• Fix RAPIDS_BRANCH version and update script (#20091) @galipremsagar
• Reduce output buffer sizes for pruned pages of columns with a list parent (#20086) @mhaseeb123
• Avoid direct CategoricalColumn calls in dask_cudf (#20080) @mroeschke
• Rework reduction case statement as dispatch_type_and_aggregation (#20078) @davidwendt
• Avoid shadowing module names (#20071) @vyasr
• Fix typing issues in pylibcudf (#20069) @vyasr
• Avoid more explicit calls to IntervalColumn and StructColumn (#20064) @mroeschke
• Cleanup of some libcudf aggregation code (#20053) @davidwendt
• Prune entries in Sphinx nitpick_ignore (#20045) @mroeschke
• Deprecate .from_pandas constructor (#19996) @mroeschke
• Improve performance of string column size computation during parquet reads. (#19986) @nvdbaranec
• Run cudf-polars conda unit tests with more than 1 process (#19980) @mroeschk...

v25.08.00

🚨 Breaking Changes

• Allow np.dtype('object') for cases that are valid (#19478) @galipremsagar
• [FEA] Remove CUDA JIT-Compatibility Checks & CCCL WARs (#19470) @lamarrr
• Drop cuda 11 usages (#19386) @galipremsagar
• Deprecate cudf::round for float types (#19298) @davidwendt
• Support output_dtype in cudf::reduce for nunique aggregation (#19265) @davidwendt
• Change default cudf-polars executor to "streaming" (#19263) @TomAugspurger
• Fix Handling of Complex Types in AST (#19248) @lamarrr
• Enable chunked reading of PQ sources with >2B rows (#19245) @mhaseeb123
• Refactor grid_1d class (#19211) @lamarrr
• Return valid for all-nulls in reduce() with nunique include-nulls aggregation (#19196) @davidwendt
• Refactor JNI error handling (#19149) @ttnghia
• Remove CUDA 11 from dependencies.yaml (#19139) @KyleFromNVIDIA
• Quick fixes of modernize-use-constraints rule (#19105) @vuule
• Filter Parquet row groups using row bounds (#19082) @mhaseeb123
• Temporarily revert "Refactor JNI error handling (#18983)" (#19076) @abellina
• Rename parquet_chunked_writer to chunked_parquet_writer for consistency with the reader (#19047) @mhaseeb123
• Compile libcudf using C++20 Standard (#19045) @vuule
• Refactor JNI error handling (#18983) @ttnghia
• stop uploading packages to downloads.rapids.ai (#18973) @jameslamb
• Remove deprecated Series methods, isclose (#18947) @mroeschke
• Remove deprecated groupby.collect (#18946) @mroeschke
• Remove deprecated get_dummies(cats=, ...) (#18944) @mroeschke
• Add pylibcudf.Column.from_arrow factory method (#18937) @Matt711
• Add pylibcudf.Table.from_arrow factory method (#18936) @Matt711
• Remove deprecated APIs (#18933) @vuule
• Remove cudf.Scalar (#18927) @mroeschke
• Remove deprecated cudf::io::host_buffer (#18881) @Matt711
• Null-handling for Transforms (#18845) @lamarrr
• Enable skip_rows in the chunked parquet reader. (#18130) @mhaseeb123

🐛 Bug Fixes

• Increase alignment requirement for parquet bloom filter to 256 (#19595) @mhaseeb123
• Revert "Add primitive row dispatch support for semi/anti join and cudf::contains" (#19503) @PointKernel
• Allow np.dtype('object') for cases that are valid (#19478) @galipremsagar
• Add conda dependency on nvidia-ml-py. (#19454) @bdice
• Mark cudf.pandas notebook repr test as flaky (#19441) @Matt711
• Fix pytest to properly expose a bug (#19433) @galipremsagar
• Switch from thrust::sort to cub::DeviceRadixSort in Parquet chunked reader (#19414) @ttnghia
• Use numba-cuda>=0.15.2,<0.16 (#19413) @bdice
• Update String Transform Examples (#19407) @lamarrr
• [BUG] Make floor division and modulo by 0 match CPU polars (#19406) @Matt711
• Handle empty input in cudf::strings::extract APIs (#19398) @davidwendt
• Fix jitify error on exit from FILTER_TEST (#19395) @davidwendt
• Update cudf.pandas tests to silence deprecation warnings (#19377) @Matt711
• Replace sprintf with snprintf in libcudf parquet tests (#19371) @davidwendt
• Make DateOffset respect timezone (#19366) @Matt711
• Fix flaky tests in cudf.pandas (#19345) @TomAugspurger
• Update protocol choices for ucxx in PDSH benchmark (#19343) @TomAugspurger
• Remove passing pandas tests from xfail list (#19341) @Matt711
• Fix Union-Slice bug (#19336) @Matt711
• Fix bit shift overflow in segmented_offset_bitmask_binop utility (#19329) @davidwendt
• Fix job filters for pandas-tests (#19322) @galipremsagar
• Fix compile warning in interop_stringview.cpp (#19320) @davidwendt
• Fix a use-after-free issue in TDigest aggregation code. (#19311) @nvdbaranec
• Always represent datetime aware data as UTC in strftime (#19304) @mroeschke
• Do not pass cupy objects objects to numba kernels directly (#19283) @brandon-b-miller
• Correct docstring for DataFrame.apply to match code (#19262) @dagardner-nv
• Cast n_unique aggregation result to match polars (#19256) @Matt711
• Fix Handling of Complex Types in AST (#19248) @lamarrr
• Add missing include (#19239) @vyasr
• Raised MixedTypeErrors for condition that lead to mixed types (#19232) @galipremsagar
• Fix errors in the nvCOMP adapter (#19221) @vuule
• Remove nvToolsExt usage (#19209) @vyasr
• Fix a pair of bugs in get_decompression_scratch() size. (#19207) @nvdbaranec
• Allow is_list_like to return correct values by disabling it (#19188) @galipremsagar
• Fix slicing after Join and GroupBy in streaming cudf-polars (#19187) @rjzamora
• Fix binops type preservation for some dtypes (#19183) @galipremsagar
• Fix streaming GroupBy on non-trivial keys (#19181) @rjzamora
• Fix bitmask in from_arrow_host for sliced stringview type (#19174) @davidwendt
• Fixed group_by mean with missing values and multiple partitions (#19165) @TomAugspurger
• Add fallback to HStack lowering in cudf-polars (#19163) @rjzamora
• Fix Literal partitioning in cudf-polars (#19160) @rjzamora
• Fix from_array_interface for empty arrays (#19144) @Matt711
• Adding GH_TOKEN pass-through to summarize job (#19143) @msarahan
• Fix hash collision in Union([MapFunction]) (#19124) @TomAugspurger
• Fix bug in group_by().n_unique() in streaming cudf-polars (#19108) @rjzamora
• Parse (non-MultiIndex) label-based keys to structured data (#19103) @mroeschke
• Fix cudf_polars spilling (#19101) @TomAugspurger
• Fix libcudf strings case logic to set null-row size to zero (#19095) @davidwendt
• Temporarily revert "Refactor JNI error handling (#18983)" (#19076) @abellina
• Temporary workaround for incorrect SplitScan results in cuDF-Polars (#19071) @rjzamora
• Use default memory resource for JSON_QUOTE_NORMALIZATION gtests (#19057) @davidwendt
• Added null-probability to polynomial benchmarks and fixed transform call-sites (#18972) @lamarrr
• Fix flaky custreamz test (#18961) @TomAugspurger
• Fix tdigest percentile correctness for low row-counts (#18952) @mythrocks
• Enable skip_rows in the chunked parquet reader. (#18130) @mhaseeb123

📖 Documentation

• Update conda environment file for CUDA 12.9 compatibility (#19376) @a-hirota
• Update recommended gcc version in contibuting guide (#19365) @davidwendt
• Autodoc DateOffset (#19297) @wence-
• Fix cudf::column_device_view::element() doxygen (#19296) @davidwendt
• Document aggregations for cudf::reduce in doxygen (#19264) @davidwendt
• add docs on CI workflow inputs (#19234) @jameslamb
• Update README and CONTRIBUTING to reflect new CUDA requirements (#19138) @PointKernel
• Remove the extra index URL for CUDA 12 (#19128) @vyasr
• Improve WordPieceVocabulary.tokenize documentation (#19098) @davidwendt
• Add some basic streaming engine documentation (#19088) @wence-
• Update the contributing guide to include pylibcudf in the build command (#19011) @Matt711
• Fix pylibcudf docs for some strings APIs (#19004) @davidwendt
• Update cuDF Python library design with BaseIndex and pylibcudf updates (#18903) @mroeschke

🚀 New Features

• Avoid using UVM on systems without a traditional memory resource (#19444) @Matt711
• Add parquet-sampling configuration options (#19423) @rjzamora
• Add new JSON reader interface accepting string column input to pylibcudf (#19400) @shrshi
• Add a parquet reader utility to update output null masks (#19370) @mhaseeb123
• Build and ship shim.cu file as LTOIR (#19368) @brandon-b-miller
• Add cudf::strings::find_instance API (#19326) @davidwendt
• Add single-file streaming Sink support (#19317) @rjzamora
• Support null_count expression (#19314) @Matt711
• Materialize tables in the experimental Parquet reader (#19308) @mhaseeb123
• Add new cudf::top_k API (#19303) @davidwendt
• Add cudf::strings::split_part API (#19289) @davidwendt
• Support output_dtype in cudf::reduce for nunique aggregation (#19265) @davidwendt
• Add post_traversal API to cudf-polars (#19258) @rjzamora
• Deprecate DataFrame.apply_rows (#19218) @brandon-b-miller
• Require numba-cuda>=0.16.0 (#19213) @brandon-b-miller
• Add a mode to co-process decompression and compression on host and device (#19203) @vuule
• Return valid for all-nulls in reduce() with nunique include-nulls aggregation (#19196) @davidwendt
• Refactor JNI error handling (#19149) @ttnghia
• Add support for horizontal string concatenation pl.concat_str (#19142) @Matt711
• Add PDS-DS Query 1 (#19131) @Matt711
• Support cudf-polars str.reverse (#19117) @brandon-b-miller
• Support cudf-polars str.pad_end and str.pad_start (#19116) @brandon-b-miller
• Support cudf-polars str.head and str.tail (#19115) @brandon-b-miller
• Support cudf-polars str.to_titlecase (#19114) @brandon-b-miller
• Add cudf/io/codec.hpp to expose compression/decompression APIs (#19113) @ttnghia
• Support converting decimals to/from pylibcudf scalars (#19106) @Matt711
• Support resource-constrained sort-merge inner join operation through left table partitioning (#19102) @shrshi
• Filter Parquet row groups using row bounds (#19082) @mhaseeb123
• Implement UDF Filters (#19070) @lamarrr
• Move the remaining libcudf pieces to C++20 (#19065) @vuule
• Allow using a stream per thread at runtime (#19051) @vyasr
• Remove stacktrace retrieval code (#19048) @ttnghia
• Compile libcudf using C++20 Standard (#19045) @vuule
• String Transform Examples: Added Branching, Public API Versions, and Sampling (#19038) @lamarrr
• Refactor JNI error handling (#18983) @ttnghia
• Add basic Sink support for streaming cudf-polars executor (#18963) @rjzamora
• Fix debug-build Failure in JIT Tests (#18939) @lamarrr
• Add from_arrow factory methods for Scalar and DataType (#18938) @Matt711
• Add pylibcudf.Column.from_arrow factory method (#18937) @Matt711
• Add pylibcudf.Table.from_arrow factory method (#18936) @Matt711
• Update nvCOMP adapter (#18931) @vuule
• Create a pylibcudf Column from a iterable of python strings (#18916) @Matt711
• Add CLI argument to enable OOM protection in PDS-H (#18914) @pentschev
• Implement data page pruning using Parquet page index stats (#18873) @mhaseeb123
• Null-handlin...