NVIDIA/cudf v26.10.00 on GitHub

What's Changed

🚨 Breaking Changes

Support zero-column operands in cross_join by @madsbk in #23235
Modify scalar factory APIs to honor the data_type scale by @davidwendt in #23483
Avoid a deep copy for no-op mergeAndSetValidity and make merging zero columns a no-op by @rishic3 in #23030
[FEA] Remove Deprecated Transform APIs by @lamarrr in #23489
Rename errc::OVERFLOW to avoid macro collisions by @a-hirota in #23509
Change cudf::strings::zfill_by_widths to accept null widths by @davidwendt in #23215
Rename cudf-polars max_io_threads configuration and change meaning by @wence- in #23569
Encode non-negative nanos for negative ORC timestamps by @vuule in #23391
Add opt-in per-row status column to VARIANT extraction and cast APIs by @abigalekim in #23560
[FEA] Simplify & Use a Meyers singleton for the cuDF context by @lamarrr in #23760
Replace the hybrid scan single-file secondary-filter API by @qbacpey in #23800
Apply memory backpressure to the shuffle by @madsbk in #23886
Fix empty mask behavior of apply_deletion_mask by @mhaseeb123 in #23857
Add Java API to return Parquet footer by @liurenjie1024 in #23912
[FEA] Drop Obsolete PTX/CUDA Rolling Aggregations by @lamarrr in #23840
Migrate stream APIs from rmm::cuda_stream_view to cuda::stream_ref by @bdice in #23929
Fix FULL_JOIN filtering through internal finalization by @bdice in #24146

🐛 Bug Fixes

Bump versions of packages missed in 26.10 rollover by @wence- in #23315
Fix NativeDepsLoaderTest is skipped in Java Test CI by @paul-aiyedun in #23312
Fix cum_sum window functions with null filling in cudf-polars by @Matt711 in #23206
Fix malformed LIB_BUILD_DIR default in conda builds by @fallintoplace in #23102
Fix 26.10 versions by @bdice in #23383
Fix hybrid scan parquet reader incorrectly parse bloom filter by @qbacpey in #22901
Make proxy extension dtypes instances of pandas ExtensionDtype by @galipremsagar in #23417
Use libcudf-bundled CCCL for NRT UDF compilation by @vyasr in #23455
Fix ORC chunked reads of empty structs by @wjxiz1992 in #23289
Fix integer overflow in nvtext::normalize_characters by @davidwendt in #23473
Fix fallback for unsupported fill strategy in cudf-polars by @rjzamora in #23485
Temporarily exclude GB300 and GH200 from testing by @KyleFromNVIDIA in #23499
Accept a device memory resource in the device-view factories by @nirandaperera in #23486
Fix usage of dask-cuda in pip constraints by @KyleFromNVIDIA in #23516
Record cudf-polars device only after a successful switch by @fallintoplace in #23512
Reject unsupported masked array-interface inputs by @fallintoplace in #23511
Fix order preservation in grouped-window evaluation by @rjzamora in #23524
[BUG] Avoid writing duplicated data in cudf-polars sink by @rjzamora in #23525
Serialize engine config in cudf-polars benchmark by @TomAugspurger in #23529
Use uint64_t to track column chunk data size in PQ writer by @mhaseeb123 in #23544
cast_variant skips unsupported target type validation for empty inputs by @abigalekim in #23462
Fix how build.sh determines CUDA version by @KyleFromNVIDIA in #23588
Fix uninitialized page info for column chunks without a dictionary page by @mhaseeb123 in #23579
Additional fixes for libcudf large strings support by @davidwendt in #23612
Fix inconsistency with Polars when doing quantile nearest on even length columns by @jberg5 in #21708
Change cudf-polars benchmarking NightlyRole date to Y-M-D format by @mroeschke in #23631
Make is_supported_cast fail for variable-width inputs by @JohnZed in #23613
Add Rocky 8 packaged-JAR Java CI and skip optional nvcomp by @paul-aiyedun in #23646
fix: declare libcudf wheel CUDA runtime dependencies by @bdice in #23414
Ensure ParquetOptions.prefetch_file_metadata value of Unspecified is serializable by @mroeschke in #23684
Guard xfail for polars test test_series_init_np_temporal_with_nat_15518 for NumPy>=2.5 by @Matt711 in #23676
Fix reset_index: level bounds, duplicate columns, columns.name, tuple index names by @galipremsagar in #23418
Fix invalid device context error in ValidIfTest.ExplicitMemoryResourcesEmptyRange by @davidwendt in #23704
Fix JSON writes to fsspec targets by @fallintoplace in #23603
Fix duplicate bin detection for large cut inputs by @fallintoplace in #23601
Require cudaMemcpyDefault policy by @bdice in #23605
Fix racecheck in parquet page_string_decode kernels by @davidwendt in #23681
Allow large strings input within the parquet make_column utility by @davidwendt in #23685
Fix host-source lifetime races by @bdice in #23718
Select batched memcpy flags by copy size by @bdice in #23675
Fix version guards in tests for set sorted hints by @Matt711 in #23703
Check for output offsets int32 overflow in list APIs by @davidwendt in #23652
Add length check to cudf::strings::is_timestamp logic for literals by @davidwendt in #23712
Reject ragged nested iterables in pylibcudf column construction by @fallintoplace in #23510
Fix per-rank result attribution in the Ray and Dask engines by @madsbk in #23739
Fix Parquet row group pruning for negated equality predicates by @mhaseeb123 in #23580
Support empty segments in segmented bitmask reductions by @vuule in #23689
Pin Cython version for cython-lint pre-commit hook by @Matt711 in #23767
Workaround cuda::std::memcpy assert in unaligned_load utility by @davidwendt in #23724
Fix StreamingOptions._from_argparse silently dropping env vars by @Matt711 in #23720
Skip upstream polars test test_scan_ndjson_streaming_decompression by @Matt711 in #23783
Fix scale handling of fixed-point types in minmax aggregation for cudf::reduce by @davidwendt in #23752
Fix RMM race conditions from deferred H2D copies in pylibcudf tests by @Matt711 in #23789
Fix flaky StreamPoolTest.ConcurrentThreadsGetDistinctStreams by @vuule in #23820
fix(to_numpy): return object dtype for bool/numeric and datetime/timedelta mixes in pandas-compatible mode by @Matt711 in #23722
Fix style check in rolling utils module by @Matt711 in #23824
Add missing include for std::cerr in debug build by @davidwendt in #23780
Fix cudf::strings::to_timestamp handling of %U format specifier by @davidwendt in #23807
Fix stream copy on scoped host variable in json reduce_to_column_tree by @davidwendt in #23806
Add max size_type overflow check in cudf::interleave_columns by @davidwendt in #23838
Remove XPASSing datetime tests from pandas XFAIL set by @Matt711 in #23866
Check for integer overflow on input parameters to cudf::tile() by @davidwendt in #23843
Add check for non-zero sample of empty table in cudf::sample() by @davidwendt in #23841
[BUG] Fix null handling in mixed semi joins by @wjxiz1992 in #23861
Handle int64 strings offsets in cudf::byte_cast API by @davidwendt in #23725
Fix fast-count optimization with hybrid scan by @TomAugspurger in #23874
Preserve header-only CSVs in pandas-compatible mode by @fallintoplace in #23513
Fix parquet prefetch test wrapper signature by @vyasr in #23883
Fix streaming builds with RAPIDS MPF stream pools by @vyasr in #23885
Fix CUDA 13 LTO architecture selection with older CMake by @wjxiz1992 in #23862
Fix Literal broadcasting in pl.struct by @TomAugspurger in #23893
Zero-fill non-nullable string offsets with a nullable ancestor by @mhaseeb123 in #23879
Fix Parquet statistics pruning for predicates satisfied by NaN by @mhaseeb123 in #23735
Require polars>=1.38 for max|min_by tests by @TomAugspurger in #23913
Fix deprecation warning: declaring default stream as cuda::stream_ref by @davidwendt in #23865
Fix segmented gather for sliced gather map input by @davidwendt in #23876
Track local row ordering in streaming cudf-polars metadata by @rjzamora in #23844
Fix compute-sanitizer racecheck hazards in DELTA_BYTE_ARRAY decode path by @pramodsatya in #23910
Fix dictionary concatenate for INT8/INT16 indices by @a-hirota in #23889
Fix stream view construction in new benchmark by @wence- in #23933
Initialize NVML once per traced do_evaluate definition by @TomAugspurger in #23926
Fix racecheck in parquet reader delta byte array kernels by @davidwendt in #23898
User upper_bound to find parquet page with end of row in get_page_span by @pmattione-nvidia in #23761
Skip failing narhwals datetime to_string tests by @TomAugspurger in #23936
Fix Parquet writer partition size computation for lists and structs by @mhaseeb123 in #23907
Avoid serializing I/O requests from Parquet reader by @mhaseeb123 in #23823
Fix build failure due to crossed merges by @davidwendt in #23954
Skip length-1 max/min_by tests with polars<1.39 by @TomAugspurger in #23969
Work around CUDA 12.9 scan corruption on Blackwell by @gerashegalov in #23973
Throw on dictionary columns in cudf::binary_operation instead of silently comparing indices by @davidwendt in #23916
Skip another test that fails due to a polars bug by @wence- in #23984
Fix dictionary specialization for json allnull_column_functor by @davidwendt in #23687
Ignore coverage of untaken branches in cudf-polars timezone handling by @wence- in #24023
Ensure that Lineariser doesn't buffer unboundedly by @wence- in #24027
Ensure streams are passed into pylibcudf calls in cudf_polars by @mroeschke in #24053
Fix racecheck reported in parquet decode_delta_length_byte_array_kernel by @davidwendt in #24034
Fix undefined behavior in Parquet varint decoding by @ttnghia in #23346
Only default pinned memory on when the system supports it by @Matt711 in #24129
Revert "Cache Cython generated sources in wheel builds (#23998)" by @wence- in #24132
Ensure CUDA Context initialized before stream sync by @TomAugspurger in #24157
Backport batched deletion-vector row count API to 26.10 by @wjxiz1992 in #24171
Backport #23919 to release/26.10 by @rishic3 in #24187
cudf doc build and CI fixes (#24279) by @TomAugspurger in #24302

📖 Documentation

Update code review guidelines to recommend spans over pointer + length parameters by @vuule in #22663
Fix Doxygen @PARAM and @tparam names that do not match the declarations by @darkdi in #23564
Update cuDF repository references for NVIDIA organization by @bdice in #23632
[DOC] Document cuDF spilling by @nethum529 in #23326
Document null binary operator semantics explicitly by @findepi in #23604
Fix stale data generation comment in generate_input.hpp by @VaggelisGian in #23816
Enable public docs features in CI by @bdice in #23902
Turn on autosectionlabel in sphinx docs by @wence- in #23942
DOC Add accelerated data engines and tools to landing page by @jsack-ga in #23921
Migrate docs links onto docs.nvidia.com by @wence- in #23971

🚀 New Features

[FEA] Add Java bindings for AST JIT operations by @thirtiseven in #23117
Use provided memory resource for wordpiece vocabulary maps by @PointKernel in #23281
Use provided memory resource for tokenize_vocabulary map by @PointKernel in #23280
Add memory resource parameter to cudf::mark_join constructors by @PointKernel in #23278
Add memory resource parameter to cudf::approx_distinct_count by @PointKernel in #23221
Add memory resource parameter to cudf::distinct_hash_join constructors by @PointKernel in #23263
Python bindings for Parquet select cols by field ID by @mhaseeb123 in #22956
Add memory resource parameter to cudf::filtered_join constructors by @PointKernel in #23264
Enable hybrid scan page pruning for page level IO by @mhaseeb123 in #23374
Add memory resource parameter to streaming_groupby by @PointKernel in #23279
Add memory resource parameter to cudf::key_remapping constructor by @PointKernel in #23277
Add floating point type support to Parquet variant field extraction by @abigalekim in #23075
Support shift over grouped windows in cudf-polars by @rjzamora in #23451
Preserve input order for grouped window evaluation by @rjzamora in #23466
Support fixed-size rolling windows inside over() by @rjzamora in #23468
Relax parquet page index requirements in hybrid scan by @mhaseeb123 in #23386
Support diff() inside over() in cudf-polars by @rjzamora in #23497
Add boolean type support to Parquet variant field extraction by @abigalekim in #23276
Add basic slicing support to pylibcudf's gpumemoryview by @Matt711 in #23541
Transcode parquet to cuDF dictionaries for flat STRING columns by @y2kiran in #22532
Support Parquet DELTA encodings with more than 64 values per mini-block by @pramodsatya in #23314
Support ordered first/last groupby aggregations in cudf-polars by @rjzamora in #23600
Add logical type inspection for raw Parquet VARIANT values by @abigalekim in #23491
Parquet Variant Extract Benchmark by @abigalekim in #23616
In-memory support for set_sorted/MapFunction("hint_sorted") by @rjzamora in #23663
Add memory resource control to valid_if by @bdice in #23490
Add GZIP compression support to the Parquet writer by @vuule in #23557
Support aggregate rolling expressions inside grouped over by @rjzamora in #23627
Update rtcx by @arhag23 in #23626
Add page-level I/O and materialization in Hybrid Scan by @mhaseeb123 in #23375
[FEA] Make JIT Compilation Warnings Only Emit in Verbose Mode by @lamarrr in #23750
Add pylibcudf bindings for fetch_byte_ranges_to_device by @Matt711 in #23543
[FEA] Implement Multi-output AST JIT & IR CSE by @lamarrr in #23621
[FEA] Add support for scalar column views in AST expressions and JIT execution by @lamarrr in #23615
Add python bindings for hybrid scan metadata and release GIL in the reader APIs by @Matt711 in #23546
Expose Parquet column-chunk statistics in pylibcudf by @rjzamora in #23666
Support streaming "hint_sorted" metadata propagation by @rjzamora in #23748
Add Arrow fixed-size-list ingress by @0guban0v in #23583
Add packaged Java tests to nightly tests by @paul-aiyedun in #23732
Add cuDF Java Maven Central publish job by @paul-aiyedun in #23714
Add Java Variant logical type inspection support by @nartal1 in #23734
Add velox-cudf CI build job by @shrshi in #22647
[FEA] Use a common base architecture as LTO IR target by @lamarrr in #23803
Collect downstream partitioning requests during cudf-polars planning by @rjzamora in #23729
Add float, boolean and array indexing support to Java Variant extraction by @nartal1 in #23711
[FEA] Introduce transform_program to amortize JIT dispatch overhead by @lamarrr in #23648
Fix kvikio thread count override and the default backend to EASY_THREADPOOL in cudf-polars by @Matt711 in #23683
Adopt hybrid scan reader in cudf-polars for split scans by @Matt711 in #23677
Extract OrderScheme metadata from a sorted Channel by @rjzamora in #22526
Optimizations for Parquet Variant's locate_object_field by @abigalekim in #23657
Use OrderScheme metadata to select order-aware Join execution by @rjzamora in #23371
Add nvtext unicode normalizer APIs by @davidwendt in #23403
Add a streaming_hash_join API for partitioned right tables by @PointKernel in #22724
[FEA] Add zstd compression support for to_csv by @a-hirota in #21518
Add decimal support to VARIANT casting by @vuule in #23858
Support dictionary columns in contiguous_split by @vuule in #23920
Add python/cython interface for unicode-normalizer APIs by @davidwendt in #23896
[MLIR] Masked Unary Ops by @brandon-b-miller in #23119
[FEA] Remove precompiled header directory specifications by @lamarrr in #23987
Support concurrent aggregate calls on one streaming_groupby by @PointKernel in #23884

🛠️ Improvements

Replace hyperlinks that 404 by @mroeschke in #23268
Replace hyperlinks that redirect by @mroeschke in #23269
Support PEP 515 underscores in string-to-integer astype by @bedo-48 in #23108
Remove redundant type dispatch from column contains by @PointKernel in #23282
[FEA] Support force_ascii flag in JSON writer by @Tjindl in #23177
Use explicit filter join kernel instantiations by @PointKernel in #23331
Reduce sort-based groupby helper build time by @PointKernel in #23285
Update RAPIDS.cmake to log source of rapids-cmake by @arhag23 in #23130
Support pl.Expr.dt.round by @mroeschke in #23158
Add assert to appropriate cudf::host_span member functions by @davidwendt in #23274
feat(cudf): add DatetimeIndex.indexer_between_time by @nethum529 in #23033
Support pl.Expr.mode/implode by @mroeschke in #23226
Support pl.Expr.dt.date by @mroeschke in #23155
Support pl.Expr.str.to_integer by @mroeschke in #23194
Support pl.Expr.str.zfill with null widths by @mroeschke in #23193
Update hooks by @vyasr in #23339
Remove redundant template instantiations from table contains by @PointKernel in #23323
Support pl.Expr.top/bottom_k_by by @mroeschke in #23213
Support pl.Expr.diff/pct_change by @mroeschke in #23225
Forward-merge release/26.08 into main by @bdice in #23382
Main merge release/26.08 by @wence- in #23395
Support pl.repeat, pl.Expr.repeat_by by @mroeschke in #23227
Fix ufunc test domains to ensure valid input generation by @Tjindl in #23196
Support pl.Expr.to_physical/hash by @mroeschke in #23313
Support pl.max_horizontal by @mroeschke in #23160
feat(pre-commit): add lychee local link checking for markdown and notebooks by @nethum529 in #23132
Label profiling thread names in libcudf and cudf-polars by @vyasr in #23408
Optimize Python overheads in Parquet metadata preprocessing by @vyasr in #23406
Forward-merge release/26.08 into main by @Matt711 in #23401
Add common load factor validation for joins by @PointKernel in #23379
Decompression - Make host fork/join stream conditional by @y2kiran in #23444
Remove redundant gather map view in sort groupby helper by @PointKernel in #23420
enforce 'yamllint' checks by @jameslamb in #23413
Update Narwhals testing with agent skill by @mroeschke in #23435
Reduce streaming groupby insert_first build time by @PointKernel in #23448
Additional Unit Tests for Parquet VARIANT Field Extraction by @abigalekim in #23036
Upgrade Thrift version to 0.23.0 by @mythrocks in #23453
Reduce filtered join build time by @PointKernel in #23320
Replace duplicate type-stringify logic with type_to_name, add print_type debug utility by @Tjindl in #23230
refactor(parquet): remove unused page_state_s fields by @vyasr in #23470
refactor(parquet): extract page_decode_setup_state substruct by @vyasr in #23471
Add cuda-profiler-api dependency for NVBench by @PointKernel in #23404
Remove deprecated sum_with_overflow APIs by @PointKernel in #23475
Reduce cudf-polars test output verbosity by @TomAugspurger in #23480
Migrate additional Thrust iterators to CUDA equivalents by @PointKernel in #23338
Add parallel native library extraction by @gerashegalov in #23409
refactor(parquet): introduce level_state_scan struct for preprocess_levels_kernel by @vyasr in #23479
Support non-unique keys for dictionary column types by @davidwendt in #22839
X-ORG-23411: Publish API docs to docs.nvidia.com by @josephine-wolf-oberholtzer in #23412
Remove dead spilling to pinned memory config option by @Matt711 in #23426
Remove deprecated key_remapping and filter_join_indices APIs by @PointKernel in #23476
Make use_cuda_wheels and cuda_suffixed explicit by @KyleFromNVIDIA in #23472
Improve Parquet reader pass construction by @mhaseeb123 in #23446
Improve performance of strings split/split_record for smaller strings by @davidwendt in #23394
Forward-merge release/26.08 into main by @gforsyth in #23531
Reduce distinct_helpers build time by @PointKernel in #23322
Migrate Thrust zip iterators to CUDA equivalents by @PointKernel in #23336
X-ORG-23411: Enable docs version picker by @josephine-wolf-oberholtzer in #23540
ensure nightly builds always produce new packages, expand 'changed-files' lists by @jameslamb in #23533
Remove bloom prefilter application in cudf-polars streaming by @wence- in #23523
Add approx_distinct_count API to cudf_streaming by @wence- in #23522
Reduce compute_shared_memory_aggs build time by @PointKernel in #23330
Remove obsolete distinct filtered join wrapper by @PointKernel in #23520
Pre-touch pages concurrently during pinned memory pool initialization by @rishic3 in #23457
[cudf] Adding memory_resources support for test utils by @nirandaperera in #23578
refactor(parquet): introduce string_offset_scan_state for preprocess_string_offsets_kernel by @vyasr in #23495
Optimize sort-merge join data passes by @PointKernel in #23532
refactor(parquet): introduce string_size_scan_state for string size kernels by @vyasr in #23496
Reduce range_rolling build time by @PointKernel in #23343
Update to rapids-logger 0.3 by @bdice in #23530
refactor(parquet): extract page_decode_nesting_state and compose full_page_decode_state replacing page_state_s by @vyasr in #23610
Reduce the number of device allocations in the ORC writer by @vuule in #23433
Update CODEOWNERS team names for rapidsai->NVIDIA migration by @josephine-wolf-oberholtzer in #23625
Optimize RLE decoding using a warp-balanced chunking approach by @vyasr in #23271
Skip Polars unit test for Numpy DeprecationWarning by @mroeschke in #23635
Use cuda::stream_ref for null mask APIs by @vyasr in #23630
Use async memory resource by default in microbenchmarks by @vuule in #23586
Use cuda::stream_ref for core libcudf APIs by @vyasr in #23645
Add annotations to pylibcudf source files by @vyasr in #23643
Change deprecated calls of stream_ref.wait() to stream_ref.sync() by @davidwendt in #23653
Support pl.Expr.str.contains/find(strict=False), pl.Expr.str.head/tail with expression inputs by @mroeschke in #23240
[cudf] Adding memory_resources support for column wrappers by @nirandaperera in #23581
Expose UCXX progress mode in cudf-polars options by @pentschev in #23659
Disable cuml-compat-tests for cpp-only PR changes by @davidwendt in #23644
wheels: enforce 'abi3audit' checks by @jameslamb in #23641
Support pl.Expr.str.to_decimal/to_integer, pl.Expr.str.count_matches/extract with more options by @mroeschke in #23243
Use cuda::stream_ref for compute APIs by @vyasr in #23649
Add numba-cuda-mlir conda package to cudf by @mroeschke in #23670
Unbind cudf::size_type from offsets used by list columns by @davidwendt in #23607
Replace <cuda/stream_ref> includes with <cuda/stream> by @bdice in #23664
Add unit tests for fixed point in quantiles by @pieroevcc in #23437
[PERF]: Improve the performance of plc.io.parquet.read_parquet with prefetched parquet file metadata. by @TomAugspurger in #23558
Change kvikio default thread count in cudf-polars to 256 by @Matt711 in #23634
Remove cudf_polars testing engine_raise_on_fail fixture by @mroeschke in #23425
Support pl.coalesce by @mroeschke in #23159
Revamp cudf-java build skill with nightly libcudf option by @rishic3 in #23248
Support pl.Expr.approx_n_unique by @mroeschke in #23211
Support pl.Expr.reverse, reverse=True for Polars cummulative expressions by @mroeschke in #23164
Use cuda::stream_ref for nested and IO APIs by @vyasr in #23669
Improve hash partition operator benchmarks by @tgujar in #23337
Parquet writer writes smaller page headers if not writing stats by @mhaseeb123 in #23431
Improve predicate pushdown coverage in cudf-polars by @Matt711 in #23143
Migrate transform iterators to CUDA by @PointKernel in #23505
Handle oversized fragments, pages, and column chunks in Parquet writer by @mhaseeb123 in #23562
Remove CUDA_STATIC_RUNTIME option by @bdice in #23698
Remove multi-file secondary-filter API and add bloom-filter pruning to hybrid scan by @qbacpey in #22861
cudf-polars benchmarks always include a query plan by @TomAugspurger in #23405
[cudf] Thread memory resources through equality/hash row preprocessing by @nirandaperera in #23665
Add more pylibcudf source annotations by @vyasr in #23673
[MINOR][cudf_polars] Validation summary skips query failures by @nirandaperera in #23508
perf(parquet): decode def and rep level streams concurrently via 2D grid by @vyasr in #23637
Use dictionary indices in join functions by @davidwendt in #23534
ci: avoid triggered-by-schedule and triggered-by-merge builds cancelling each other by @jameslamb in #23716
Ignore xgboost warning in cudf.pandas third party tests by @mroeschke in #23650
Reduce libcudf binary size by trimming instantiations by @vyasr in #23706
Remove pip only skips of numba-cuda-mlir backend tests by @brandon-b-miller in #23713
Change LIST column offsets from size_type to int32 in benchmarks and gtests by @davidwendt in #23682
Reduce Parquet writer memory footprint for list columns by @mhaseeb123 in #23573
Use OrderScheme metadata to select order-aware groupby execution by @rjzamora in #23306
Disable benchmarks in cudf-polars test runner, add cudf-polars dependency on kvikio by @bdice in #23719
Add parquet benchmark for output_dict_columns option by @y2kiran in #23596
Optimize hash full join by @PointKernel in #23521
Use no:xdist on pylibcudf stream checking pytests by @davidwendt in #23721
Add Ordering.as_strict helper by @rjzamora in #23757
Make the CUDA stream pool per-thread by @vuule in #23672
Remove Arrow C++ dependency from C++ tests by @vyasr in #23708
Use cuda::stream_ref in remaining core libcudf APIs by @vyasr in #23691
Binary search optimization for Parquet Variant's object field ID lookup by @abigalekim in #23638
X-ORG-23411: Publish dask-cudf to docs.nvidia.com by @josephine-wolf-oberholtzer in #23728
Use cuda::stream_ref in Python stream shims by @vyasr in #23694
Migrate to cuco's key-based Bloom filter policy by @PointKernel in #23671
Use cuda::stream_ref in Java native code by @vyasr in #23726
refactor: replace rmm::device_scalar with cudf::detail::device_scalar by @vyasr in #23618
Fix dictionary pruning for column chunks with no dictionary page by @pmattione-nvidia in #23746
Report kvikio I/O statistics per rank in the streaming engines by @madsbk in #23738
Support pl.Expr.hist/is_close by @mroeschke in #23254
Support pl.Expr.entropy by @mroeschke in #23384
Add --sleep-between-iterations option to cudf-polars benchmark runner by @Matt711 in #23763
fix(update-version): make numba-cuda upper pin creation idempotent by @gforsyth in #23815
Support groupby.mean() for decimal columns; fix decimal reflection binops in cudf Python by @mroeschke in #23787
Hybrid scan avoids nullable cols unless pruning pages by @mhaseeb123 in #23792
Use new constraints type from DFG by @KyleFromNVIDIA in #23556
[MINOR] Route leftover temps through local temp_mr by @nirandaperera in #23730
Enable numpydoc validation by @pathak-satyam in #23439
Replace rolling.apply implementation with numba-cuda-mlir by @mroeschke in #23598
Support pyarrow 24 by @vyasr in #23809
Refactor Parquet synthetic column helpers by @mhaseeb123 in #23759
Capture engine setup/teardown durations by @TomAugspurger in #23808
Add pylibcudf source annotations for stubgen by @vyasr in #23715
PERF: Optimize cudf-polars stable hashing by @TomAugspurger in #23582
Support pl.Expr.kurtosis/skew by @mroeschke in #23353
pre-commit: update 'cython-lint' to 0.21.1 by @jameslamb in #23842
Migrate T-digest test helpers to memory_resources by @bdice in #23608
Use cuda::stream_ref in libcudf tests by @vyasr in #23768
Migrate Nanoarrow test helpers to memory_resources by @bdice in #23609
Add nvcarps skill by @vyasr in #23851
Use cuda::stream_ref in benchmarks and examples by @vyasr in #23769
Shard pandas tests across runners in PR and nightly CI by @galipremsagar in #22992
Add PDSHDuckDBQueries for cudf.pandas PDSH benchmarks; support validation with decimal data by @mroeschke in #23788
Deprecate windowslinetermination CSV option by @utkarshparekh in #23363
Speed up segmented_bitmask_and for tall tables by @vuule in #23686
Add parquet DELTA read tests and nested benchmark axis by @vyasr in #23771
CI: narrow cudf.pandas unit test filter by @vyasr in #23869
Update skill with latest validation by @vyasr in #23852
Enable Sphinx version switcher for dask-cudf docs by @josephine-wolf-oberholtzer in #23870
Add apply_retention_mask and deprecate apply_boolean_mask by @mhaseeb123 in #23700
Support Series/DataFrame.mean with decimals; allow cudf.pandas PDSH benchmarks to run with fallback-as-error by @mroeschke in #23871
Use cuda::stream_ref in libcudf_streaming by @vyasr in #23727
Use different max_concurrent_io_tasks defaults for local and remote paths by @rjzamora in #23847
Decimal cleanup by @TomAugspurger in #23518
Replace hand-rolled output iterators with cuda::tabulate_output_iterator by @PointKernel in #23856
Test one Polars version per matrix entry in cudf-polars wheel tests by @Matt711 in #23867
Lazily initialize streaming groupby aggregation kinds by @bdice in #23850
Reuse sampled parquet footers from statistics collection by @Matt711 in #23855
Let callers reserve partitioning memory by @madsbk in #23833
wheels: declare 'cuda-pathfinder' dependency for libcudf by @jameslamb in #23880
CI: Migrate custom jobs to RTX PRO 6000 by @bdice in #23878
[MLIR] MaskedType: numeric binary operators by @brandon-b-miller in #22886
Add compressed percentile approximation oracle test by @vyasr in #23798
Support pl.Expr.max_by/min_by by @mroeschke in #23224
Improve Parquet writer struct test by @mhaseeb123 in #23882
Avoid blocking memory reservations in actor coroutines by @madsbk in #23892
Improve strings split on whitespace performance for smaller strings by @davidwendt in #23542
Make Parquet statistics pruning null-aware by @pmattione-nvidia in #23747
Fix racecheck in segmented_offset_bitmask_binop kernel by @davidwendt in #23897
Expose prepend_source_index_column parquet reader options in pylibcudf by @mroeschke in #23911
Support pl.Expr.dt.convert_time_zone/replace_time_zone by @mroeschke in #23424
Optimize DICT32 transcode via batched keys and index shift by @y2kiran in #23710
Use cuda::stream_ref for central stream helpers by @vyasr in #23770
Convert cudf-polars ActorTracer to a dataclass by @TomAugspurger in #23925
Add cudf-polars DataFrame._size_bytes by @TomAugspurger in #23924
Update cudf-polars Quent export format by @TomAugspurger in #23922
CI: Reduce dask-cuDF wheel test frequency by @vyasr in #23941
Give distinct exit codes to query and validation errors in benchmark runner by @mroeschke in #23928
CI: Remove redundant Conda Java test jobs by @vyasr in #23946
CI: Preserve notebook test trigger by @vyasr in #23947
Route dictionary encode temp allocations through memory_resources by @nirandaperera in #23642
CI: Enable sccache preprocessor cache for Java builds by @vyasr in #23943
CI: Filter CMake tests and C++ linters by source changes by @vyasr in #23945
Support pl.Expr.mean_horizontal/min_horizontal/sum_horizontal by @mroeschke in #23427
Update codeowners for narwhals test plugins by @TomAugspurger in #23940
Add a JNI wrapper for reusable distinct hash join by @rishic3 in #22392
Run narwhals tests on cudf-polars changes by @mroeschke in #23961
Improve cudf-spark-jni build workflow by @bdice in #23825
Fix jit compile warning for CUDA 12.9 by @davidwendt in #23944
Update hybrid scan benchmark I/O policy by @bdice in #23957
CI: Narrow conda C++ test triggers by @vyasr in #23962
CI: Narrow Java test triggers by @vyasr in #23950
CI: Narrow docs build triggers by @vyasr in #23964
Update cudf-polars Quent attribute serialization by @TomAugspurger in #23951
Move Velox compatibility build to nightly tests by @bdice in #23948
CI: Consolidate Python test triggers by @vyasr in #23968
Support Polars Array dtype pass-through in cudf-polars by @0guban0v in #23773
CI: Narrow C++ file triggers by @vyasr in #23970
Unpin pandas<3.0.4a as 3.0.4 was yanked and patched in 3.0.5 by @sethtroisi in #23577
CI: Narrow Python package test triggers by @vyasr in #23975
Copy many cudf-polars benchmark runner functionality over to cudf.pandas benchmark runner by @mroeschke in #23979
CI: Refine PR changed-file filters by @vyasr in #23976
Replace hash join internals with HashCSR by @PointKernel in #23640
Use cuDF to count the total deleted row count for cuDF based Delta reader for deletion vectors by @sdrp713 in #23978
CI: Shard cudf-polars compatibility tests by @vyasr in #23990
CI: Use isolated build for all wheel packages by @vyasr in #23991
CI: Stage wheel builds by package type by @vyasr in #23988
Cache Cython generated sources in wheel builds by @vyasr in #23998
Build noarch packages in parallel by @vyasr in #23997
Consolidate cuDF wheel tests by @vyasr in #23994
CI: Reduce cached Conda C++ build overhead by @vyasr in #23999
CI: Restore cudf-spark-jni sccache setup by @vyasr in #23996
CI: Report slow Python tests by @vyasr in #24009
TEST: Remove redundant concat tests by @vyasr in #24011
TEST: Reuse I/O test fixtures by @vyasr in #24012
TEST: Reuse invariant Python test inputs by @vyasr in #24013
TEST: Speed up GroupBy test setup by @vyasr in #24014
TEST: Merge import subprocess checks by @vyasr in #24010
TEST: Remove deterministic temporal and rolling skips by @vyasr in #24015
TEST: Remove deterministic reshape skips by @vyasr in #24016
Java CI: remove CUDA version normalization and install ninja via pip by @paul-aiyedun in #23790
Adopt CUDA stream compatibility accessors by @bdice in #23995
Update dependencies after dep-file change by @wence- in #24057
Improve Ordering reuse for groupby and join by @rjzamora in #23927
Rework join filter pushdown rewrites to leave hints and apply adaptively by @wence- in #23584
Remove the -Wno-error=deprecated-declarations compile flag by @davidwendt in #23985
Default pinned host memory to enabled for cudf-polars' Ray, Dask, and SPMD engines by @Matt711 in #23836
Default remote parquet metadata work to off by @pentschev in #24111
Default cudf-polars to the kvikio MULTI_POLL backend by @Matt711 in #23839
Improve and rework metadata handling in the hybrid scan reader by @mhaseeb123 in #23795
Pin Polars<1.45 by @Matt711 in #23914
Rewrite skills docs links by @wence- in #24083
Fix UnicodeNormalizer from_python_unicodedata performance by @davidwendt in #24063
Use CUDA driver API for context initialization in pylibcudf by @bdice in #24151
Expose filter_join_indices in Java by @bdice in #24097
X-ORG-410: Integrate archived docs into version switcher by @josephine-wolf-oberholtzer in #24398

New Contributors

@bedo-48 made their first contribution in #23108
@arhag23 made their first contribution in #23130
@darkdi made their first contribution in #23564
@pieroevcc made their first contribution in #23437
@pathak-satyam made their first contribution in #23439
@VaggelisGian made their first contribution in #23816
@sethtroisi made their first contribution in #23577

Full Changelog: v26.10.00a...v26.10.00