v26.04.00

What's Changed

🚨 Breaking Changes

• Enable Explicit Context Shutdown by @madsbk in #794
• [Refactor] AllowOverbooking enum by @madsbk in #797
• String utilities by @madsbk in #803
• Move to_device coroutine to TableChunk by @wence- in #832
• Overhaul/Cleanup of Errors by @madsbk in #831
• Removing cudax components by @nirandaperera in #856
• Remove multi-message support in shuffle chunks by @madsbk in #863
• Rename Node → Actor by @madsbk in #866
• Statistics cleanup by @madsbk in #871
• Reimplement allreduce using a recursive doubling scheme by @wence- in #872
• Introduce Statistics::record_copy and consolidate copy/spill statistics by @madsbk in #875
• Remove logger from progress thread by @wence- in #889
• Stream ordered timings by @madsbk in #882
• Split the streaming context and communicator apart by @wence- in #895

🐛 Bug Fixes

• Make C++ SPDX headers consistent by @pentschev in #798
• Pass objects safely from Python coroutines to C++ coroutines by @pentschev in #796
• Fix option rename: memory_reserve_timeout_ms => memory_reserve_timeout by @madsbk in #815
• Revert microseconds representation to "us" by @pentschev in #823
• read_parquet Node bug by @nirandaperera in #824
• Add missing wait to StreamingMemoryReserveOrWait.CheckPriority by @pentschev in #836
• Raise missing std::invalid_argument by @pentschev in #837
• Avoid using "p2p" shuffle in dask-cudf test by @rjzamora in #877
• Promote nested closures in run_actor_network to top-level by @wence- in #878
• Add ucxx as a runtime dependency by @Matt711 in #876
• Fix deprecated commandline args for mpirun in ctest launch by @wence- in #886
• Update RMM imports for cuda_stream deprecation by @TomAugspurger in #888

📖 Documentation

• Add a glossary to the documentation by @TomAugspurger in #759
• Reorganize the docs by @madsbk in #870
• fix links in docs by @madsbk in #879

🚀 New Features

• Add AllReduce collective by @pentschev in #683
• Introduce standard ChannelMetadata definition by @rjzamora in #819
• Add Slurm support to rrun with PMIx-based coordination by @pentschev in #775
• Add more bootstrap utils to Python API by @pentschev in #868
• Add "contiguous" partition-assignment variation by @rjzamora in #891

🛠️ Improvements

• Option: dask_spill_to_pinned_memory by @madsbk in #800
• Python bindings for MemoryReserveOrWait by @madsbk in #787
• One more C++ string utils by @madsbk in #813
• Update versions in pre-commit config by @wence- in #812
• Config options: add typed insert_if_absent<T> by @madsbk in #816
• Report "means" in bench_comm by @pentschev in #820
• Add mechanism for sending metadata through channels by @wence- in #811
• Update partition logic to accommodate cudf changes by @PointKernel in #805
• Split ndsh validation into run and validate by @TomAugspurger in #801
• Add from_options() factory methods by @madsbk in #825
• Introduce the coroutine reserve_memory() and use it in the ndsh workflows by @madsbk in #827
• Handle date types in ndsh queries by @TomAugspurger in #799
• Support PackedData from host bytes in Python by @rjzamora in #828
• Increase tag space for OpID by @wence- in #826
• Drop Python 3.10 support by @gforsyth in #829
• Enable setting/ updating cuda stream in Buffer by @nirandaperera in #814
• tighten wheel size limits, expand CI-skipping logic, other small build changes by @jameslamb in #830
• Use TopologyDiscovery from cuCascade by @pentschev in #774
• Migrate bloom filter and related utilities into rapidsmpf proper by @wence- in #834
• Rename columns to column_names in parquet reader options by @Matt711 in #835
• remove pip.conf migration code in CI scripts, update CI-skipping rules by @jameslamb in #839
• New make_table_chunks_available_or_wait() by @madsbk in #838
• Add python bindings for streaming bloom filter by @wence- in #841
• Removing packed_columns from TableChunk by @nirandaperera in #840
• More string parsing utils by @madsbk in #845
• Use verify-hardcoded-version pre-commit hook by @KyleFromNVIDIA in #846
• Add unbounded per-context file read cache by @madsbk in #842
• Minor performance fixes by @wence- in #849
• Use GHA id-token for sccache-dist auth token by @trxcllnt in #852
• Streaming Q4 implementation by @TomAugspurger in #710
• Update clang-format's include order by @madsbk in #861
• Rename columns to column_names in q04 by @Matt711 in #860
• Expose local_partitions of shuffler to Python by @wence- in #862
• Add safe_cast utility for overflow-safe integer conversions by @madsbk in #864
• Style Inconsistency Fixes by @madsbk in #867
• refactor: build wheels and conda packages using Python limited API by @gforsyth in #850
• Statistics: decouple formatters from stats by @madsbk in #880
• refactor(limited api): add explicit wheel.py-api to pyproject.toml
  by @gforsyth in #887
• Statistics: add JSON export by @madsbk in #883
• Use read_parquet_metadata to get column types by @Matt711 in #865
• Let the communicator own a progress thread by @wence- in #884
• Simplify python setup of communicators and progress thread in integrations by @wence- in #892
• Store logger as a shared pointer by @wence- in #894
• Update Cython lower bound pin to 3.2.2 by @vyasr in #893
• Remove pytest upper bound pin by @vyasr in #898
• Add support for Python 3.14 by @gforsyth in #881
• Install rrun by @madsbk in #900
• Add configurable pool properties to PinnedMemoryResource by @nirandaperera in #851
• Add ability to obtain the shape of a TableChunk independently of where the packed data resides by @wence- in #905
• Add devcontainers output to dependencies.yaml that excludes keys that pull in ray-default by @trxcllnt in #906
• Rename rrun environment variables prefix to RRUN_ by @pentschev in #908
• test_gather_shuffle_statistics(): only check for existence by @madsbk in #912
• Clarify and ensure op_id reuse after local extraction is safe for AllGather and AllReduce by @wence- in #909
• Remove ack phase in shuffler communication event loop by @wence- in #910
• build wheels with CUDA 13.0.x, test wheels against mix of CTK versions by @jameslamb in #919
• docs-build: use Python ...

v26.02.00

What's Changed

🚨 Breaking Changes

• TableChunk.copy(): allow wiggle room by @madsbk in #672
• Reorganize buffer/memory code by @madsbk in #673
• Reorganize ScopedMemoryRecord by @madsbk in #676
• Don't build pip devcontainer by @wence- in #670
• Expose async collectives to Python by @wence- in #685
• Memory type refactor by @madsbk in #701
• Host buffer by @madsbk in #696
• PinnedMemoryResource refactor by @madsbk in #714
• PinnedMemoryResource: use unlimited release_threshold by @madsbk in #735
• Introducing MemoryType::PINNED_HOST by @madsbk in #731
• Python bindings for PinnedMemoryResource by @madsbk in #742
• Refactor a Context's coro::thread_pool by @madsbk in #776
• Context now owns a MemoryReserveOrWait instance by @madsbk in #783
• Backport #794: Enable Explicit Context Shutdown by @pentschev in #809

🐛 Bug Fixes

• Disable sccache for clang-tidy builds by @pentschev in #695
• Add rapids-generate-version by @bdice in #699
• Fix data races by @madsbk in #694
• Add missing memory type parameters docs by @pentschev in #703
• Fix read parquet multi rank by @wence- in #708
• Fix retrieving number of ranks when running benchmarks with rrun by @pentschev in #709
• Install OpenMPI explicitly in Python test matrix by @pentschev in #720
• Add missing nogil for shuffler insertion by @rjzamora in #722
• Pin Ray<2.52 by @pentschev in #730
• Re-enabling linking all tests by @nirandaperera in #741
• Force BUILD_TESTING flag for CTest by @bdice in #739
• q09: fix cudf::transform() call. by @madsbk in #746
• Fix rrun's --bind-to options by @pentschev in #749
• Fix finding of preexisting rapidsmpf build (for environments like devcontainers) by @vyasr in #754
• [MINOR] Fix build.sh by @nirandaperera in #756
• Fix non_device_size calculation in unpack_and_concat by @TomAugspurger in #765
• Trigger CI by @TomAugspurger in #768
• Remove mode argument from to_pylibcudf call by @pentschev in #770

📖 Documentation

🚀 New Features

• Add missing rrun support to C++ benchmarks by @pentschev in #679
• Add missing rrun support to Python benchmark by @pentschev in #680
• Support rrun binding to resources by @pentschev in #704

🛠️ Improvements

• [MINOR] Increase CI timeout to 10m by @nirandaperera in #665
• Use strict priority in CI conda tests by @bdice in #684
• Use rapidsai/sccache fork in conda-cpp-linters job by @trxcllnt in #697
• Python bindings for Context.spillable_messages() by @madsbk in #690
• bench-shuffle: discard output by @madsbk in #689
• Adding Fanout node by @nirandaperera in #636
• Use strict priority in CI conda tests by @bdice in #700
• TPCH-derived Q9 by @wence- in #663
• HostMemoryResource: Enable Transparent Huge Pages (THP) by @madsbk in #712
• SpillableMessages: support message copy by @madsbk in #713
• Remove alpha specs from non-RAPIDS dependencies by @bdice in #715
• SPILL_TARGET_MEMORY_TYPES by @madsbk in #718
• Extend bench_memory_resources by @madsbk in #719
• Enable merge barriers by @KyleFromNVIDIA in #721
• Formalize Buffer Memory Types and Storage Mapping by @madsbk in #729
• Add static library for cudf_tests by @pentschev in #733
• Enable clang-tidy checks in build.sh by @pentschev in #728
• Include spill amounts in NVTX payloads by @TomAugspurger in #717
• Add devcontainer fallback for C++ test location by @bdice in #734
• Reorder bench_shuffle barriers by @pentschev in #744
• libcoro: use the newest commit by @madsbk in #747
• Make unbounded fanout messages spillable by @nirandaperera in #711
• Implement Q1 and Q3 and refactor to provide utilities by @wence- in #738
• Compatibility with CCCL 3.2 by @bdice in #755
• CI: run bench_memory_resources smoke tests by @madsbk in #748
• ndsh: add --no-pinned-host-memory by @madsbk in #752
• Cudf pack/ chunked_pack benchmark by @nirandaperera in #745
• Empty commit to trigger a build by @bdice in #763
• [MINOR] Fix build warnings by @nirandaperera in #762
• Use SPDX license identifiers in pyproject.toml, bump build dependency floors by @jameslamb in #766
• Add CUDA 13.1 support by @bdice in #750
• build and test against CUDA 13.1.0 by @jameslamb in #767
• Allow CMake v4+ by @trxcllnt in #669
• Test pip devcontainers in CI by @trxcllnt in #727
• Host buffer with a type-erased owner by @nirandaperera in #764
• Memory reserve or wait by @madsbk in #688
• System info utils by @madsbk in #743
• Cython coroutine handling by @madsbk in #769
• Add streaming benchmark validator by @TomAugspurger in #760
• Empty commit to trigger a build by @jameslamb in #782
• Basic streaming read_parquet benchmark by @wence- in #753
• Multi-GPU implementation of Q21 by @wence- in #758
• Use main shared-workflows branch by @jameslamb in #784
• Add device PCI Bus ID information to check_resource_binding by @pentschev in #785
• wheel builds: react to changes in pip's handling of build constraints by @mmccarty in #790
• fix(build): build package on merge to release/* branch by @gforsyth in #821

New Contributors

• @mmccarty made their first contribution in #790

Full Changelog: v26.02.00a...v26.02.00

v25.12.00

What's Changed

🚨 Breaking Changes

• Use a CUDA stream pool and remove many stream parameters from the API by @madsbk in #531
• Add optional owner slot to TableChunks by @wence- in #536
• Clean up ShufflerAsync by @madsbk in #557
• Deallocation should be noexcept by @bdice in #569
• Table chunk is_spillable() by @madsbk in #559
• Upgrade libcoro dependency by @wence- in #570
• Refactor: Message by @madsbk in #598
• TableChunk copying by @madsbk in #604
• Promote sequence IDs to messages by @wence- in #602
• Message with content description by @madsbk in #623
• Refactor: Context.create_channel() by @madsbk in #631
• Context: use shared pointer for buffer resource by @madsbk in #646
• [MINOR] Adding a Range arg to BufferResource::reservce_and_fail by @nirandaperera in #648
• Implement linearisation by round-robin assignment by @wence- in #654
• Handle lifetime issues in channel callbacks in presence of exceptions by @wence- in #661

🐛 Bug Fixes

• Enable timeout with stacks for ctests by @pentschev in #538
• tests: use initialized input by @madsbk in #541
• Attempt to address deadlocks in streaming shuffle tests by @wence- in #548
• [MINOR] Fix python build by @nirandaperera in #562
• Use a semaphore in async shuffler to manage waiters by @wence- in #564
• Replace path to user's home with wildcard by @pentschev in #579
• Update to Ray 2.49 by @pentschev in #585
• Release the CuptiMonitor mutex before joining thread by @pentschev in #588
• Fix hang when a consumer throws by @madsbk in #577
• Table.from_table_view_of_arbitrary: Fix missing stream argument. by @madsbk in #607
• Fix a few SPDX-related issues by @KyleFromNVIDIA in #614
• Ensure messages are dropped by refcount through channel sends by @wence- in #638
• AllGather unregisters spill function in destructor by @madsbk in #643
• Pin Cython pre-3.2.0 and PyTest pre-9 by @jakirkham in #647
• refactored update-version.sh to handle new branching strategy by @rockhowse in #645

📖 Documentation

• Fix wording in background doc by @TomAugspurger in #568
• Capitalize GPU by @bdice in #603
• Docs for streaming engine by @quasiben in #571
• Fix doxygen warnings and amend docstrings by @quasiben in #606
• Use current system architecture in conda environment creation command by @bdice in #634
• Fix images by @quasiben in #677

🚀 New Features

• Make UCXX progress mode configurable by @pentschev in #487
• Add an arbitrary payload chunk by @wence- in #620
• Add RapidsMPF launcher rrun by @pentschev in #616
• Separate shuffler's communication into new interface by @pentschev in #437
• Support for dynamic system topology discovery by @pentschev in #624

🛠️ Improvements

• Update RAPIDS_BRANCH, codify changes in update-version.sh by @KyleFromNVIDIA in #529
• Make CI scripts more consistent and trap exit error codes by @pentschev in #542
• Fix hang in boostrap_dask_cluster with new clients by @TomAugspurger in #540
• Compute sanitizer suppressions by @madsbk in #544
• Fixing forward merge conflict by @nirandaperera in #558
• Separate synced host data and buffers by @nirandaperera in #543
• ShutdownAtExit: accept a vector of channels by @madsbk in #563
• Add an async interface to the AllGather implementation by @wence- in #551
• Fix rmm.librmm cimport by @bdice in #575
• Use new remove feature from UCXX tagProbe by @pentschev in #581
• Enable sccache-dist connection pool by @trxcllnt in #586
• Lock-free pausable thread loop by @nirandaperera in #583
• Adding pinned host buffer impl by @nirandaperera in #549
• Use main in RAPIDS_BRANCH by @bdice in #591
• Use main shared-workflows branch by @bdice in #595
• Coroutine utility function: coro_results() by @madsbk in #597
• Enable taking Python Streams from Context by @nirandaperera in #596
• [MINOR] Remove rmm/host/pinned_memory_resource by @nirandaperera in #605
• Use SPDX for all copyright headers by @KyleFromNVIDIA in #608
• Memory Reservation python bindings by @madsbk in #610
• Message copy by @madsbk in #609
• Introduce a streaming read_parquet node by @wence- in #574
• Content description of an object by @madsbk in #613
• Minor fixes by @wence- in #622
• Use a large stack size when running lineariser test by @madsbk in #625
• Minor type stub fixes by @wence- in #626
• Make CUDA stream pool size configurable by @TomAugspurger in #633
• Migrate to new CCCL memory resource interface by @bdice in #635
• Collection for spillable messages by @madsbk in #630
• Update some Python docstrings by @madsbk in #639
• More python bindings by @madsbk in #651
• Some test renaming by @madsbk in #660
• Spilling of in-flight messages by @madsbk in #640
• Add statistics to BufferResource wrapper by @TomAugspurger in #659
• Update RMM includes from <rmm/mr/device/*> to <rmm/mr/*> by @bdice in #664
• Improved serialization of Options by @pentschev in #642
• [MINOR] Adding a RAPIDSMPF_DETAIL mode by @nirandaperera in #612
• Use sccache-dist build cluster for conda and wheel builds by @trxcllnt in #628
• Install OpenMPI explicitly in CI matrices by @pentschev in #723

New Contributors

• @quasiben made their first contribution in #571
• @rockhowse made their first contribution in #645

Full Changelog: v25.12.00a...v25.12.00