v26.06.00

What's Changed

🚨 Breaking Changes

• Simplify shuffle completion tracking by @wence- in #914
• Simplify async shuffle completion notification by @wence- in #916
• Implement streaming allreduce and use it in streaming bloom filter construction by @wence- in #928
• Switch to std::string_view where possible in Backend by @pentschev in #951
• Split out utilities in allgather and simplify completion tracking by @wence- in #957
• Execute PyActors in asyncio.TaskGroup by @mroeschke in #943
• Make BufferResource parameters required instead of optional by @vyasr in #961
• Replace cudaMemcpyAsync with rapidsmpf::cuda_memcpy_async by @nirandaperera in #965
• Report failures programmatically in rrun.bind() by @pentschev in #971
• Predefine statistics formatters by @madsbk in #976
• Migrate RMM usage to CCCL MR design by @bdice in #940
• Add serialize/deserialize to Python Statistics by @madsbk in #980
• Change cuCascade GIT_TAG to main by @nirandaperera in #988
• Refactor memory resource ownership and type safety by @nirandaperera in #985
• Remove memory resources from Statistics by @nirandaperera in #1003
• Disable pinned memory by default by @nirandaperera in #1020
• Remove py_executor from run_actor_network by @wence- in #1023
• Remove Dask integration by @madsbk in #1049

🐛 Bug Fixes

• [MINOR] Fixing weakref on None by @nirandaperera in #917
• Fix UCXX import for pip installations by @madsbk in #921
• Fix allgather op_id reuse better by @wence- in #918
• Support reuse of op_ids in shuffles by @wence- in #927
• Fix Lineariser::drain() deadlock by @Matt711 in #932
• Avoid defining awaitable in PyActor.__init__ by @mroeschke in #941
• Add individual timeout for each of ctest nranks run by @pentschev in #945
• Move away from deprecated ucxx::Address::getString() by @pentschev in #953
• Forward channel metadata through bloom filter apply by @wence- in #952
• Increase timeout for test_bootstrap_multiple_clients by @pentschev in #962
• Use non-deprecated cudf::filtered_join constructor by @pentschev in #973
• Fix ReceivedChunks::spill() throwing on control messages by @Matt711 in #975
• More BufferResource lifetime fixes by @vyasr in #970
• Fix ReceivedChunks::spill() over-spilling past the amount target by @Matt711 in #984
• Suppress warnings from cuco headers, do not export CMake target by @nirandaperera in #990
• Remove unnecessary $<BUILD_INTERFACE:...> from CMake targets by @pentschev in #1001
• Fix PostBox::spill() passing stale amount to std::ranges::lower_bound by @Matt711 in #983
• Fix librapidsmpf wheel builds with current cuDF packages by @pentschev in #1022
• Remove C language from librapidsmpf wheel builder by @pentschev in #1024
• Skip stack trace if gdb is unavailable by @pentschev in #1004
• Fix large-message initialization in communication benchmark by @pentschev in #1016
• Use new cuCascade::cucascade_topology_discovery target by @KyleFromNVIDIA in #1030
• Remove unnecessary include from slurm_backend.cpp by @pentschev in #1034
• Fix socket backend test write handling by @bdice in #1033
• More follow-ups to recent cucascade fixes by @KyleFromNVIDIA in #1035
• Remove CUDA includes from bootstrap sources by @bdice in #1044
• Pin cuCascade for 26.06 release by @pentschev in #1057

📖 Documentation

• Fix formatting and language issues in docs by @pentschev in #929

🚀 New Features

• Move resource binding to callable bind() function by @pentschev in #949
• Python bindings for rrun.bind() by @pentschev in #950
• Add functions for resource binding verification to librrun by @pentschev in #972
• Add skill to reproduce CI locally by @pentschev in #991
• Enable Ray for aarch64 with conda by @pentschev in #1002
• Add TableChunk.into_packed_data by @Matt711 in #986
• Add socket-based bootstrap coordination backend for rrun by @pentschev in #1000
• [C++] Add OrderScheme partitioning metadata by @rjzamora in #993
• Add hybrid Slurm support to rrun with PMIx-based coordination by @pentschev in #844
• [Python] Add OrderScheme partitioning metadata by @rjzamora in #853
• Add support for Ray in Python 3.14 and wheels tests by @pentschev in #1015

🛠️ Improvements

• Avoid fork method in dask tests by @TomAugspurger in #901
• Fix compiler warnings for deprecated APIs by @Matt711 in #903
• Forward-merge release/26.04 into main by @jameslamb in #923
• [MINOR] Adding python bindings for reserve_or_fail by @nirandaperera in #925
• Simplify Shuffler postboxes by @wence- in #920
• Dask cluster bootstrap with options and enable pinned memory by default by @nirandaperera in #926
• Limit pinned reservations by @nirandaperera in #935
• update pip devcontainers' base image tags by @trxcllnt in #937
• [MINOR] Improve error message for duplicate option keys by @nirandaperera in #938
• Pre-distribute listener addresses during UCXX barrier() by @pentschev in #934
• Remove deprecated benchmark::internal::Benchmark usage by @bdice in #939
• [MINOR] Verbose nvtx traces by @nirandaperera in #936
• Report pinned memory resource profile stats by @nirandaperera in #948
• Update to clang 20.1.8 by @bdice in #956
• Add BufferResource lifetime tracking to Python wrapper classes by @vyasr in #960
• Gather statistics by @madsbk in #958
• Integrate MetadataPayloadExchange into Shuffler by @pentschev in #649
• [MINOR] Use cudf::packed_size to estimate the memory usage of a table. by @nirandaperera in #967
• Migrate from pynvml to cuda.core.system by @mdboom in #969
• Add to_dict() to Python Statistics by @madsbk in #989
• Implement a sparse alltoall exchange pattern by @wence- in #959
• Harden FileBackend key validation and temp file creation by @pentschev in #992
• fix(ci): resolve all zizmor findings and add zizmor pre-commit checks by @gforsyth in #1007
• Add sql text validation to ndsh.py by @TomAugspurger in #1006
• Use token.rapids.nvidia.com when issuing S3 bucket creds in devcontainers by @trxcllnt in #1005
• Topology-aware default pool sizing for PinnedMemoryResource by @nirandaperera in #1012
• Simplify HostBuffer memory ownership by @nirandaperera in #1013
• Build and test with CUDA 13.2.0 by @bdice in #1021
• Use static linkage for CUDA runtime by @bdice in #725
• skip CuPy 14.1.0 by @jameslamb in #1068

New Contributors

• @mdboom made their first contribution in #969

Full Changelog: v26.06.00a...v26.06.00