NVIDIA/cudf v26.04.00 on GitHub

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 #21575
Add stream parameter to indexalator make_input_optional_iterator by @davidwendt in #21545
Adapt to latest rapidsmpf API changes by @wence- in #21601
Usage of template keyword required for some nvcc and gcc combinations by @robertmaynard in #21577
Fix caching the result of is_sorted by @jberg5 in #21603
Small bugfix: json sink logic was being evaluated incorrectly by @jberg5 in #21605
fix: correct typos in comments by @thecaptain789 in #21380
Fix chunked_pack::next() returning overflowed int for >2GB data by @ttnghia in #21643
Adjust precision in pds-h query 11 by @TomAugspurger in #21651
Fix segfault in HYBRID_SCAN_MULTITHREAD_NVBENCH by @ttnghia in #21661
Improve attribute lookup by 500x in cudf.pandas by @galipremsagar in #21656
Fix pl.Expr.split_exact children result for null data by @mroeschke in #21660
Increase timeout for the devcontainer build workflow by @davidwendt in #21713
Support large partition counts in hash_partition by @ttnghia in #21674
Use larger datatypes when calculating mean and other groupby-aggs by @quasiben in #21719
Remove default stream use in PQ reader by @mhaseeb123 in #21629
Fix dynamic-join logic for minimum shuffle modulus by @rjzamora in #21712
Support fixed point types in Parquet predicate pushdown by @mhaseeb123 in #21681
Fix inherited partitioning metadata in cudf-polars by @rjzamora in #21748
Fix validation failures in PDS-DS Q23 and Q61 by @Matt711 in #21754
Fix validation failures in Q69 and Q93 by @Matt711 in #21758
Fix thread pool worker threads using wrong CUDA device in multi-GPU environments by @vuule in #21745
Fix quantile on decimals by @jberg5 in #21707
Add missing return after state.skip() in libcudf benchmarks by @davidwendt in #21756
[BUG] Fix parsing of ld.param PTX instruction in UDFs by @lamarrr in #21597
Enable astype(str) in cudf by @galipremsagar in #21699
Fix racecheck in join mark_retrieve_kernel by @davidwendt in #21786
Fix approx distinct count nan handling by @PointKernel in #21806
Fix mypy pinning on rmm by @bdice in #21925
Fix JIT transform kernel crash on non-default CUDA streams by @bdice in #21921

📖 Documentation

cudf -> pd typo in README by @jrbourbeau in #21134
Fix docs for cudf::index_of_first_set_bit API by @davidwendt in #21313
[DOC] Document Java ColumnView reductions [skip ci] by @rishic3 in #21205
Remove outdated issue links by @rishic3 in #21259
Move Parquet IO utils to a new doxygen group by @mhaseeb123 in #21412
Remove link to deprecated JIT-unspilling documentation by @TomAugspurger in #21710

🚀 New Features

Add remaining pandas PDSH queries by @Matt711 in #21108
Add the fast type to the NamedAgg wrapper in cudf.pandas by @Matt711 in #21135
ENH: Expose split_part to Python API via pylibcudf by @Umang-projects in #21068
Experimental deletion vector parquet APIs to accept information via const ref to a struct by @mhaseeb123 in #21044
Add cudf::index_of_first_set_bit API by @davidwendt in #21207
Expose cudf::filter to python by @Matt711 in #21300
Add a Java wrapper for deletion vector APIs by @jihoonson in #21231
Add allgather_reduce utility for the rapidsmpf runtime by @rjzamora in #21339
Add Hybrid scan APIs to chunked materialize all columns by @mhaseeb123 in #21377
Util to get packed table buffer size by @nirandaperera in #21322
Add PDS-DS Q11 by @Matt711 in #19658
Move hybrid scan IO utils to libcudf by @mhaseeb123 in #21360
Expose column_nans_to_nulls API to Python and update cudf to use it by @vyasr in #21438
Add PDS-DS Q12-16 by @Matt711 in #20389
[FEA] Enable Pre-compiled Headers for faster JIT by @lamarrr in #21457
Support named capture groups in cudf extract by @davidwendt in #21484
Use cuda::std::optional for parquet LogicalType to enable device compatibility by @PointKernel in #21470
[FEA] update cudf::transform and cudf::filter callsites by @lamarrr in #21537
Example to demonstrate inter-parquet-file pipelining using hybrid scan APIs by @mhaseeb123 in #20722
Add PDS-DS Q30-36 by @Matt711 in #20393
[FEA] Fix deprecation warnings in cudf::transform and cudf::filter by @lamarrr in #21551
Retrieve cudf::data_type schema from Parquet metadata by @Matt711 in #21458
Add divmod support, respect array priority in ufuncs, and properly handle negative infs and zeros by @vyasr in #21523
Add PDS-DS Q63-72 by @Matt711 in #20400
Dynamic Distinct and GroupBy support by @rjzamora in #21433
Support ArrowStringView from device in cudf::from_arrow_device by @davidwendt in #21385
Add PDS-DS Q24-29 by @Matt711 in #20392
Add set_decimal_width option to parquet reader by @Matt711 in #21472
Add PDS-DS Q17-23 by @Matt711 in #20390
Add PDS-DS Q73-85 by @Matt711 in #20402
Add PDS-DS Q86-99 by @Matt711 in #20404
Add PDS-DS Q37-43 by @Matt711 in #20394
[FEA] Introduce udf_source_type enum to distinguish between CUDA and PTX UDFs. by @lamarrr in #21567
Add PDS-DS Q53-62 by @Matt711 in #20399
Add PDS-DS Q44-52 by @Matt711 in #20395
Benchmark parquet footer read and column select time by @mhaseeb123 in #21422
Return file footer buffer at Parquet writer close by @mhaseeb123 in #21378
Add dynamic join actor by @rjzamora in #21578
Add --io-mode for page cache control (cold/lukewarm/hot) by @Matt711 in #21653
Expand and improve AST expression support in Parquet readers by @mhaseeb123 in #21515
Support count aggregation (valid/all) for cudf::scan by @davidwendt in #21565
[FEA] Add JIT cache management functions by @lamarrr in #21553
Add mark_join to support semi/anti join with left table reuse by @shrshi in #21440
Restore multithreaded optimization in the CSV reader by @vuule in #21307

🛠️ Improvements

Add no_racecheck filter to snap, unsnap, and inflate kernel names by @davidwendt in #21043
Main merge release/26.02 by @mmccarty in #21071
Disable unneeded checking in QUANTILES_TEST during racecheck runs by @davidwendt in #21011
Optionally write results from pdsh benchmarks by @TomAugspurger in #21052
Main merge release/26.02 by @vyasr in #21073
Use CRTP (Curiously Recurring Template Pattern) to implement clone() method for aggregation classes by @ttnghia in #21066
Add a hybrid scan API to build an all-true row mask by @aryansri05 in #20862
[polars-rapidsmpf] Shutdown RapidsMPF's Context by @madsbk in #21083
Allow for scheduler file and existing dask cluster when using pdsh by @quasiben in #21024
Remove children from ColumnBase by @vyasr in #21077
Add new pinned vector factory functions by @mhaseeb123 in #21033
Add std::span operator to cudf span classes by @davidwendt in #21053
[polars-rapidsmpf] enable spill to pinned memory by @madsbk in #21067
Expand Metadata for better multi-GPU planning information by @rjzamora in #20908
Removal IntervalDtype/StructDtype inheritance by @vyasr in #21114
install 'binutils' in conda Python test env by @jameslamb in #21165
Add missing stream synchronizations to various tests by @vyasr in #21122
Skip setitem PeriodDtype tests by @Matt711 in #21164
Refactor groupby aggregation, removing simple_aggregations_collector and aggregation_finalizer classes by @ttnghia in #21064
Use verify-hardcoded-version pre-commit hook by @KyleFromNVIDIA in #21152
Use cub FlaggedIf for copy-if with stencil parameter by @davidwendt in #21156
Fix pl.Expr.fill_null in cudf_polars with strategy "zero"/"one" for boolean columns by @mroeschke in #21146
Use Span thoughout pylibcudf by @Matt711 in #21087
Ignore hardcoded versions in deprecated code by @bdice in #21192
Replace thrust::count_if with calls to cudf::detail::count_if by @davidwendt in #21162
Add cudf::detail::reduce_by_key_async utility by @davidwendt in #21184
Fix for DeviceMin and DeviceMax default device operators by @davidwendt in #21118
Forward merge release/26.02 into main by @Matt711 in #21191
Move pylibcudf tests to correct directory by @vyasr in #21188
Use the same channel for TableChunk and Metadata messages by @rjzamora in #21182
Add new ColumnBase.create API by @vyasr in #21187
Use ColumnBase.create more by @vyasr in #21216
tighten wheel size limits, expand CI-skipping logic, other small build changes by @jameslamb in #21220
Remove numba-cuda upper bound by @brandon-b-miller in #21117
Drop Python 3.10 support by @gforsyth in #21225
Added an extra-info field to pds-h output by @TomAugspurger in #21219
Allow IR nodes to use multiple collective IDs by @rjzamora in #21203
Remove public scan/reduce ColumnBase APIs by @vyasr in #21217
Remove pandas compatibility gating in cudf.factorize by @galipremsagar in #21238
Remove pandas compatibility gating for dtypes in join by @galipremsagar in #21239
Remove pandas compatibility gating in frame and groupby methods by @galipremsagar in #21254
Add DynamicPlanningOptions configuration by @rjzamora in #21250
Remove pandas compatibility gating of dtypes in Index methods by @galipremsagar in #21255
Update verify-copyright configuration by @bdice in #21196
Remove pandas compatibility gating for dtypes in timdelta column by @galipremsagar in #21258
Move rapids-logger statement inside if in test_cpp.sh by @davidwendt in #21266
Reduces CategoricalColumn reliance on codes and improve property cache clearing by @vyasr in #21264
Remove pandas compatibility gating for dtypes in struct, list and interval columns by @galipremsagar in #21240
Use ColumnBase.create more in StringColumn by @mroeschke in #21265
Replace thrust::copy_if calls with cudf::detail::copy_if by @davidwendt in #21268
Update rolling logic to allow unsorted dictionary keys by @davidwendt in #21037
Remove deprecated pylibcudf arrow APIs by @Matt711 in #21284
Don't gate ColumnBase.to_pandas behind mode.pandas_compatible by @mroeschke in #21256
Remove pandas compatibility gating for dtypes in DataFrame APIs by @galipremsagar in #21261
Remove pandas compatibility gating for dtypes in decimal column by @galipremsagar in #21243
Remove pandas compatibility gating for dtypes in rolling window by @galipremsagar in #21260
remove pip.conf migration code in CI scripts, update CI-skipping rules by @jameslamb in #21303
Don't hard-code branch names by @KyleFromNVIDIA in #21270
Change CUB If and FlaggedIf to use cudf::detail::copy_if by @davidwendt in #21274
Migrate to rapidsmpf ChannelMetadata by @rjzamora in #21275
Remove deprecation warnings from benchmarks/ndsh/utilities.cpp by @davidwendt in #21308
Improve pylibcudf usage in cudf by @vyasr in #21285
Continue refactoring column construction by @vyasr in #21262
Is already partitioned by @rjzamora in #21310
Consolidate cudf-classic dtype validators to their own file and make them stricter by @mroeschke in #21281
Preprocess parquet repetition and definition levels by @pmattione-nvidia in #21139
Remove pandas compatibility dtypes gating in series methods by @galipremsagar in #21257
Implement na for match and contains in str API by @galipremsagar in #21236
Remove redundant lowering and dispatching infrastructure for the rapidsmpf runtime by @rjzamora in #21337
Full structlog-based tracing by @rjzamora in #21316
Remove some unneeded libcudf debug-specific code paths by @davidwendt in #21335
Move has_nulls template parameter to runtime in rolling window by @davidwendt in #21312
Enable KvikIO datasource in hybrid scan examples by @mhaseeb123 in #21318
Reserve more collective IDs for "dynamic" IR nodes by @rjzamora in #21343
Deprecate values_host in favor of to_numpy by @Matt711 in #21346
refactor: build wheels and conda packages using Python limited API by @gforsyth in #21329
Ensure ColumnBase.create asserts that plc_column type is equivalent to passed dtype by @mroeschke in #21325
Move generate_nested_types.hpp definitions to .cpp file by @davidwendt in #21344
Remove dead code checking scheduler in fallback logic by @Matt711 in #21351
Skip failing cuDF Python tests on RTX PRO 6000 by @bdice in #21358
Remove mode.pandas_compatible from StringColumn.all_characters_of_type by @mroeschke in #21356
Remove export from cudf::detail::sort and cudf::detail::merge APIs by @davidwendt in #21345
Cleanup duplicated code in hybrid scan by @mhaseeb123 in #21353
Make custreamz and dask-cudf independent of cuda-version by @vyasr in #21348
Use ColumnBase.create more in cudf classic APIs by @mroeschke in #21306
Avoid single-partition assumptions when dynamic-planning is enabled by @rjzamora in #21370
Add a fallback-mode cli arg to the pdsh benchmarks by @Matt711 in #21365
Add benchmarks for dictionary set_keys and sort by @davidwendt in #21366
fix(deps): bump pyarrow floor in conda recipes by @gforsyth in #21383
More create usage and tightening up dtype specification by @vyasr in #21359
Optimize cudf usage of pylibcudf by @vyasr in #21362
Include a run_id in pds-h benchmark run by @TomAugspurger in #21375
Capture query plan in pds-h output by @TomAugspurger in #21315
Remove cudf compat gating at more places by @galipremsagar in #21350
Use ColumnBase.create more in frame.py by @mroeschke in #21394
Remove ._with_dtype_metadata in cudf classic groupby by @mroeschke in #21389
Allocate single compressed buffer per source in parquet reader by @vuule in #21323
Preserve ArrowDtype in Series.struct.fields by @mroeschke in #21379
Use ColumnBase.create instead of from_pylibcudf in indexed_frame.py by @mroeschke in #21401
Use GHA id-token for sccache-dist auth token by @trxcllnt in #21406
Remove more mode.pandas_compatible from StringColumn by @mroeschke in #21393
Rewrite CategoricalColumn operations in terms of pylibcudf and avoid _with_type_metadata by @vyasr in #21402
Remove from_pylibcudf/_with_dtype_metadata in IO readers/rolling by @mroeschke in #21409
Remove Frame._copy_type_metadata method by @vyasr in #21405
Apply PIMPL idiom to sort-merge join by @shrshi in #21349
Avoid _with_dtype_metadata in dask_cudf by @mroeschke in #21411
Remove hash_vocab_utils by @vyasr in #21419
Refactor find_and_replace in Numerical and String columns by @vyasr in #21413
Use ColumnBase.create over .from_pylibcudf in MultiIndex, groupby, misc. by @mroeschke in #21423
Skip RTX PRO 6000 tests in test workflow by @vyasr in #21431
Split up algorithm.cuh into reduce.cuh and copy_if.cuh by @davidwendt in #21387
Put typename disambiguator before nested name specifier by @seanbaxter in #21408
Use ColumnBase.create over .from_pylibcudf in dataframe/index/indexed_frame.py by @mroeschke in #21439
Remove mode.pandas_compatable from ColumnBase.cast by @mroeschke in #21441
Remove mode.pandas_compatible from ColumnBase._validate_args by @mroeschke in #21426
Make rapidsmpf py_executor thread pool size configurable by @TomAugspurger in #21446
Replace thrust::constant_iterator with cuda::constant_iterator by @miscco in #21445
Remove _with_type_metadata by @vyasr in #21421
Drop thrust::reverse_iterator in favor of cuda::std::reverse_iterator by @miscco in #21455
Move detail::sorted_order decl from sort_column_impl.cuh to sort.hpp by @davidwendt in #21436
Remove remaining uses of ColumnBase.from_pylibcudf by @mroeschke in #21448
Add public bitmask_and with host_span parameters by @davidwendt in #21437
RapidsMPF renamed Node to Actor by @madsbk in #21476
Reduce nvbench parameter space for high-multiplicity and sort-merge join benchmarks by @shrshi in #21076
Reduce PR workflow jobs for cpp-only changes by @davidwendt in #21381
Optimize PDSH queries for cudf.pandas by @galipremsagar in #21462
Pin polars >=1.30 and <1.39 by @Matt711 in #21304
Use param substitutions from dsqgen by @Matt711 in #21465
Make cudf-polars config generic over its executor by @TomAugspurger in #21459
Allow running PDS-DS with qualification parameters by @Matt711 in #21478
Use new local_partitions method of shuffler by @wence- in #21461
Unify normalization, buffer wrapping, and validation by @vyasr in #21451
Update/remove version checks for outdated versions in Python libraries by @mroeschke in #21483
check-nightly-ci: update to new version by @jameslamb in #21464
Remove pyarrow string_view workarounds in pylibcudf/cudf by @mroeschke in #21490
Remove dead decimal alignment code by @Matt711 in #21474
Move Column normalization to only required code paths by @vyasr in #21487
Rename columns to column_names in parquet reader options by @Matt711 in #21473
Fix more deprecated calls in JNI source by @davidwendt in #21482
Remove inefficiencies in gather by @galipremsagar in #21429
pds-h validation improvements by @TomAugspurger in #21388
check-nightly-ci: remove testing config by @jameslamb in #21494
Update RapidsMPF’s Context() since it no longer takes a statistics argument. by @madsbk in #21503
Drop uses of thrust/functional by @miscco in #21477
Refactor some common utilities out of find_and_replace by @vyasr in #21500
Sort specialization logic for dictionary column by @davidwendt in #21420
Add cudf::detail::copy_if_async utility by @davidwendt in #21495
Update range values for gather/scatter benchmarks by @davidwendt in #21479
Replace some cudf::detail calls with public API equivalents by @davidwendt in #21507
Reenable pandas tests disabled due to prior CoW incompatibilities by @vyasr in #21510
Add utility to extract partitioning moduli from the ChannelMetadata by @rjzamora in #21491
Refresh cudf.pandas xpass list by @vyasr in #21521
List and struct support for make_empty_column by @quasiben in #21493
Create ColumnBase.from_range to reduce as_column usage by @mroeschke in #21516
Add properties in serialized explain for Scan and Shuffle by @TomAugspurger in #21497
Implement sink nodes for rapidsmpf runtime by @TomAugspurger in #20965
Reduce CI by skipping tests unaffected by changes based on package dependency chain by @Matt711 in #21506
Remove unneeded specialized device_vector/device_span gtest by @davidwendt in #21508
Rework reductions/scan_tests.cpp to help with adding aggregation types by @davidwendt in #21511
Fix pdsh benchmark help text by @TomAugspurger in #21524
Update casts in pds-h for --validate-directory by @TomAugspurger in #21534
Change thrust::make_discard_iterator to cuda::make_discard_iterator by @davidwendt in #21526
Avoid cudf.dtype dtype parsing internally by @mroeschke in #21522
Use --rmm-release-threshold for both single- and multi-gpu benchmarks by @rjzamora in #21535
Test cudf-polars with the "rapidsmpf" runtime and "distributed" cluster by @rjzamora in #21028
chore(deps): bump numba ceiling to <0.65.0 by @gforsyth in #21543
Support JIT filtering of join output indices by @shrshi in #21202
Use stricter dtype object checks in cudf classic by @mroeschke in #21527
Use new modules for RMM CudaStreamFlags by @bdice in #21559
chore(deps): bump numba upper bound in conda recipe by @gforsyth in #21561
Validate PDS-DS Queries 1-16 by @Matt711 in #21556
Update conftest-patch for minor numpy changes by @vyasr in #21568
Create wrapper for pylibcudf function calls by @vyasr in #21501
Set track-stream-ordered-races=no in compute-sanitizer racecheck for LARGE_STRINGS_TEST by @davidwendt in #21562
Skip scan decimal introspection logic when there are no decimals in the schema by @Matt711 in #21571
Add per-file, unprivileged page cache dropping to cuDF benchmarks by @kingcrimsontianyu in #21397
Use non-blocking streams in libcudf's internal stream pool by @vuule in #21449
Make cudf-polars pds-h script more robust by @TomAugspurger in #21538
Remove Index materialization in all query results by @galipremsagar in #21542
refactor(limited api): add explicit wheel.py-api to pyproject.toml
by @gforsyth in #21560
Fix make_empty_histogram_like to call make_empty_column by @davidwendt in #21558
Make IntervalColumn more agnostic to pandas extension types by @mroeschke in #21589
Use UTC time more in DatetimeTZColumn by @mroeschke in #21566
Add validation options to cudf.pandas PDSH runner by @mroeschke in #21569
Avoid ArrowDtype conversion in ColumnBase.create by @mroeschke in #21528
Save duckdb "gold" results to a directory without validating by @Matt711 in #21599
Add null_count() back to cudf::column_device_view by @davidwendt in #21430
Make Struct/ListColumn respect copy(deep=True) by @mroeschke in #21583
Use plc.interop.ColumnMetadata for more ColumnBase.to_arrow conversions by @mroeschke in #21586
Avoid using pyarrow.Scalar as intermediary in cuDF by @mroeschke in #21591
Add hybrid scan multithreaded benchmark by @mhaseeb123 in #21363
Pass timeout-minutes to devcontainer CI job by @mroeschke in #21610
Update Cython lower bound pin to 3.2.2 by @vyasr in #21612
feat: add support for Python 3.14 by @gforsyth in #21540
Pin upper bound of numba-cuda to <0.28.0 by @mroeschke in #21628
Handle float edge cases in pds-h validation by @TomAugspurger in #21585
Add pylibcudf interface for cudf::make_empty_lists_column by @davidwendt in #21564
Validate PDS-DS Queries 63-72 by @Matt711 in #21588
Pass plc.interop.ColumnMetadata to arrow conversion in ColumnBase.element_indexing by @mroeschke in #21587
Adjust retried exceptions in fetch-versions by @TomAugspurger in #20954
Use PylibcudfFunction in more places by @vyasr in #21611
Add cudf::is_dictionary_key() traits utilities by @davidwendt in #21574
chore: followup fixes for Python 3.14 and test filtering by @gforsyth in #21622
Preserve partitioning metadata after simple Select operations by @rjzamora in #21547
Validate PDS-DS Queries 17-23 by @Matt711 in #21594
Make Struct/ListColumn more agnostic to pandas extension types by @mroeschke in #21550
Align JIT filter join naming with libcudf convention by @shrshi in #21621
Validate PDS-DS Queries 24-29 by @Matt711 in #21592
Add shift and shift_fill support by @quasiben in #21662
include pytorch conda packages in CUDA 13 test env by @jameslamb in #21663
Update defaults for the "rapidsmpf" runtime by @rjzamora in #21668
Remove validation from normalization by @vyasr in #21664
Minor follow-up cleanups from merged PRs by @vyasr in #21658
Introduce packed_metadata_view utility for introspecting cudf::packs metadata wire format by @wence- in #21670
Reduce parameters for JSON_READER_NVBENCH to improve setup runtime by @davidwendt in #21666
Split up PARQUET_READER_NVBENCH benchmark by @davidwendt in #21667
Improvement for libcudf benchmark utility create_structs_data by @davidwendt in #21657
Validate PDS-DS Q30-36 by @Matt711 in #21576
Enable dynamic-planning by default with the "rapidsmpf" runtime by @rjzamora in #21675
Use sccache and sccache-dist in spark-rapids-jni workflow by @trxcllnt in #21682
Validate PDS-DS Queries 37-43 by @Matt711 in #21615
Unpin numba-cuda upper bound and adjust groupby.apply fallback conditons by @brandon-b-miller in #21637
Validate PDS-DS Queries 44-48 by @Matt711 in #21634
Validate PDS-DS Queries 49-52 by @Matt711 in #21636
Validate PDS-DS Queries 53-57 by @Matt711 in #21633
Use PylibcudfFunction for simple table functions by @vyasr in #21678
Validate PDS-DS Queries 58-62 by @Matt711 in #21640
Validate PDS-DS Queries 78-82 by @Matt711 in #21648
Replace max_precision() utility with STL function by @davidwendt in #21691
Change detail calls of mask functions to public equivalent by @davidwendt in #21563
Use cudaMemcpyBatchAsync by @bdice in #20800
Remove nvidia-libnvcomp from dependencies.yaml by @trxcllnt in #21679
Validate PDS-DS Queries 73-77 by @Matt711 in #21647
Validate PDS-DS Queries 86-92 by @Matt711 in #21617
Embed PDS-DS query parameters as Python module instead of JSON file by @Matt711 in #21715
Validate PDS-DS Queries 93-99 by @Matt711 in #21618
Remove unused dictionary::detail::get_insert_index utility by @davidwendt in #21711
Use CUDF_EXPECTS when the condition has no side effects by @vuule in #21722
Validate PDS-DS Queries 83-85 by @Matt711 in #21649
Remove libcudf gtest detail_gather_test.cu by @davidwendt in #21735
Move true_if device function from detail/.hpp to src/.cuh header by @davidwendt in #21717
RapidsMPF SPMD mode in PDSH benchmarks by @madsbk in #21652
Remove benchmark of benchmark utility create_structs_data by @davidwendt in #21724
Use ThreadPoolExecutor contextmanager to ensure graceful shutdown in cudf_polars with rapidsmpf by @mroeschke in #21753
Remove unneeded internal make_strings_column factory by @davidwendt in #21747
Use CUB topK as fast-path for cudf::top_k by @davidwendt in #21582
fix verify-hardcoded-versions issues by @jameslamb in #21770
Cap numba-cuda upper bound at <0.29.0 by @brandon-b-miller in #21785
enforce a floor on libnvjitlink, build wheels with CUDA 13.0.x, test wheels against mix of CTK versions by @jameslamb in #21671

New Contributors

@jrbourbeau made their first contribution in #21134
@Umang-projects made their first contribution in #21068
@drivanov made their first contribution in #21400
@seanbaxter made their first contribution in #21408
@thecaptain789 made their first contribution in #21380

Full Changelog: v26.04.00a...v26.04.00