[Core] fix heap corruption in RayletClient causing driver crash (SIGABRT: "corrupted size vs. prev_size")#58660
Merged
edoakes merged 1 commit intoray-project:masterfrom Nov 18, 2025
Conversation
Contributor
There was a problem hiding this comment.
Code Review
This pull request correctly identifies and fixes a critical use-after-free bug in RayletClient by using std::shared_from_this to extend the object's lifetime for asynchronous RPC callbacks. The changes in both the header and implementation file are appropriate for resolving the heap corruption issue. I have one suggestion to improve the robustness of the fix by reordering two lines to prevent a potential inconsistency in the pins_in_flight_ counter, as was also noted in the pull request description.
b264cd8 to
b8e5e43
Compare
b8e5e43 to
2be315b
Compare
2be315b to
a36c9a2
Compare
| /// [RayletClientInterface] and works on worker registration, lease management, etc. | ||
| class RayletClient : public RayletClientInterface { | ||
| class RayletClient : public RayletClientInterface, | ||
| public std::enable_shared_from_this<RayletClient> { |
There was a problem hiding this comment.
Bug: Ensure shared_from_this compatibility with stack objects.
Adding std::enable_shared_from_this<RayletClient> breaks compatibility with stack-allocated RayletClient instances. The codebase has at least one stack-allocated instance at src/ray/core_worker/core_worker_process.cc:887. If PinObjectIDs is called on such an instance, shared_from_this() will throw std::bad_weak_ptr, causing a crash. While the current stack-allocated instance only calls GetSystemConfig, this creates a latent bug where future code changes could inadvertently call PinObjectIDs on stack-allocated instances.
Collaborator
There was a problem hiding this comment.
I'll fix this as a follow up
a36c9a2 to
d597abf
Compare
…BRT: "corrupted size vs. prev_size") Signed-off-by: dragongu <andrewgu@vip.qq.com>
d597abf to
c1b8584
Compare
edoakes
reviewed
Nov 17, 2025
Collaborator
edoakes
left a comment
There was a problem hiding this comment.
Thanks for the detailed report and fix @dragongu! Did you check to see if there are any other response callbacks that suffer from the same issue?
An alternative solution to this would be to ensure that we always stop the io_context prior to destructing the RayletClient. IMO that would be a bit cleaner of a solution, as I don't like the idea of proliferating our use of shared_from_this (it can be used as a crutch to cover up sloppy coding practices). We are already doing it in the graceful path, but not if the core worker is destructed directly. Did you consider this?
Contributor
Author
@edoakes Yes, stopping by priority can solve it, but relying on external calls for the correctness of the callback doesn’t seem very ideal？ Also, has this issue already been indirectly fixed? |
edoakes
approved these changes
Nov 18, 2025
| /// [RayletClientInterface] and works on worker registration, lease management, etc. | ||
| class RayletClient : public RayletClientInterface { | ||
| class RayletClient : public RayletClientInterface, | ||
| public std::enable_shared_from_this<RayletClient> { |
Collaborator
There was a problem hiding this comment.
I'll fix this as a follow up
400Ping
pushed a commit
to 400Ping/ray
that referenced
this pull request
Nov 18, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 ray-project#1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 ray-project#2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 ray-project#3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 ray-project#4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 ray-project#5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 ray-project#7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 ray-project#8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 ray-project#9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 ray-project#10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 ray-project#11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 ray-project#12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 ray-project#13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 ray-project#14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 ray-project#19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 ray-project#20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 ray-project#25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 ray-project#26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 ray-project#28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 ray-project#30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 ray-project#31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 ray-project#32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 ray-project#33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 ray-project#35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 ray-project#36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 ray-project#37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 ray-project#38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 ray-project#1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 ray-project#2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 ray-project#3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 ray-project#4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 ray-project#5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 ray-project#6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 ray-project#7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 ray-project#8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 ray-project#9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 ray-project#10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 ray-project#11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 ray-project#12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 ray-project#13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 ray-project#14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 ray-project#15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 ray-project#16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 ray-project#17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 ray-project#18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 ray-project#23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com>
edoakes
pushed a commit
to edoakes/ray
that referenced
this pull request
Nov 18, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 #1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 #2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 #3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 #4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 #5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 #7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 #8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 #9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 #10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 #11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 #12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 #13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 #14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 #19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 #20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 #25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 #26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 #27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 #28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 #29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 #30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 #31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 #32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 #33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 #35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 #36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 #37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 #38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 #1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 #2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 #3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 #4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 #5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 #6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 #7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 #8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 #9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 #10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 #11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 #12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 #13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 #14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 #15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 #16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 #17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 #18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 #23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com>
Aydin-ab
pushed a commit
to Aydin-ab/ray-aydin
that referenced
this pull request
Nov 19, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 #1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 ray-project#2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 ray-project#3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 ray-project#4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 ray-project#5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 ray-project#7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 ray-project#8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 ray-project#9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 ray-project#10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 ray-project#11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 ray-project#12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 ray-project#13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 ray-project#14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 ray-project#19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 ray-project#20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 ray-project#25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 ray-project#26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 ray-project#28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 ray-project#30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 ray-project#31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 ray-project#32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 ray-project#33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 ray-project#35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 ray-project#36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 ray-project#37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 ray-project#38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 #1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 ray-project#2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 ray-project#3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 ray-project#4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 ray-project#5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 ray-project#6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 ray-project#7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 ray-project#8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 ray-project#9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 ray-project#10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 ray-project#11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 ray-project#12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 ray-project#13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 ray-project#14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 ray-project#15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 ray-project#16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 ray-project#17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 ray-project#18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 ray-project#23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com> Signed-off-by: Aydin Abiar <aydin@anyscale.com>
ykdojo
pushed a commit
to ykdojo/ray
that referenced
this pull request
Nov 27, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 ray-project#1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 ray-project#2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 ray-project#3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 ray-project#4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 ray-project#5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 ray-project#7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 ray-project#8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 ray-project#9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 ray-project#10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 ray-project#11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 ray-project#12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 ray-project#13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 ray-project#14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 ray-project#19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 ray-project#20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 ray-project#25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 ray-project#26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 ray-project#28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 ray-project#30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 ray-project#31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 ray-project#32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 ray-project#33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 ray-project#35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 ray-project#36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 ray-project#37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 ray-project#38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 ray-project#1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 ray-project#2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 ray-project#3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 ray-project#4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 ray-project#5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 ray-project#6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 ray-project#7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 ray-project#8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 ray-project#9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 ray-project#10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 ray-project#11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 ray-project#12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 ray-project#13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 ray-project#14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 ray-project#15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 ray-project#16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 ray-project#17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 ray-project#18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 ray-project#23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com> Signed-off-by: YK <1811651+ykdojo@users.noreply.github.com>
SheldonTsen
pushed a commit
to SheldonTsen/ray
that referenced
this pull request
Dec 1, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 ray-project#1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 ray-project#2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 ray-project#3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 ray-project#4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 ray-project#5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 ray-project#7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 ray-project#8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 ray-project#9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 ray-project#10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 ray-project#11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 ray-project#12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 ray-project#13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 ray-project#14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 ray-project#19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 ray-project#20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 ray-project#25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 ray-project#26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 ray-project#28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 ray-project#30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 ray-project#31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 ray-project#32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 ray-project#33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 ray-project#35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 ray-project#36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 ray-project#37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 ray-project#38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 ray-project#1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 ray-project#2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 ray-project#3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 ray-project#4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 ray-project#5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 ray-project#6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 ray-project#7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 ray-project#8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 ray-project#9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 ray-project#10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 ray-project#11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 ray-project#12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 ray-project#13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 ray-project#14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 ray-project#15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 ray-project#16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 ray-project#17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 ray-project#18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 ray-project#23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com>
Future-Outlier
pushed a commit
to Future-Outlier/ray
that referenced
this pull request
Dec 7, 2025
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 #1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 #2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 #3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 #4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 #5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 #7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 #8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 #9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 #10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 #11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 #12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 #13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 #14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 #19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 #20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 #25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 #26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 #27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 #28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 #29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 #30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 #31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 #32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 #33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 #35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 #36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 #37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 #38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 #1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 #2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 #3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 #4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 #5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 #6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 #7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 #8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 #9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 #10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 #11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 #12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 #13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 #14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 #15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 #16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 #17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 #18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 #19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 #20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 #21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 #22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 #23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com> Signed-off-by: Future-Outlier <eric901201@gmail.com>
peterxcli
pushed a commit
to peterxcli/ray
that referenced
this pull request
Feb 25, 2026
…BRT: "corrupted size vs. prev_size") (ray-project#58660) ## Summary This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the `RayletClient` object is destroyed while an asynchronous RPC callback is still pending. ## Problem Description ### Scenario A Ray Data job (Ray 2.50.0) with pipeline `read_parquet -> filter -> map_batches -> write` running for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT: ``` corrupted size vs. prev_size *** SIGABRT received at time=1761916578 on cpu 30 *** PC: @ 0x7f073569d9fc (unknown) pthread_kill Aborted (core dumped) ``` ### Trigger Conditions After reproducing with an ASan image, Asan reveals the actual use-after-free at: ``` #0 0x7ff282967361 in std::__atomic_base<long>::fetch_sub(long, std::memory_order) /usr/include/c++/11/bits/atomic_base.h:628 ray-project#1 0x7ff282967361 in std::__atomic_base<long>::operator--(int) /usr/include/c++/11/bits/atomic_base.h:377 ray-project#2 0x7ff282967361 in operator() src/ray/raylet_rpc_client/raylet_client.cc:338 ray-project#3 0x7ff282967361 in __invoke_impl<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:61 ray-project#4 0x7ff282967361 in __invoke_r<void, ray::rpc::RayletClient::PinObjectIDs(const ray::rpc::Address&, const std::vector<ray::ObjectID>&, const ray::ObjectID&, ray::rpc::ClientCallback<ray::rpc::PinObjectIDsReply>&)::<lambda(ray::Status, ray::rpc::PinObjectIDsReply&&)>&, const ray::Status&, ray::rpc::PinObjectIDsReply> /usr/include/c++/11/bits/invoke.h:111 ray-project#5 0x7ff282967361 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#6 0x7ff2829fbadf in std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>::operator()(ray::Status const&, ray::rpc::PinObjectIDsReply&&) const /usr/include/c++/11/bits/std_function.h:590 ray-project#7 0x7ff2829fbadf in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}::operator()(ray::Status const&) const bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:293 ray-project#8 0x7ff2829fbadf in void std::__invoke_impl<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(std::__invoke_other, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:61 ray-project#9 0x7ff2829fbadf in std::enable_if<is_invocable_r_v<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>, void>::type std::__invoke_r<void, ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status>(ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}&, ray::Status&&) /usr/include/c++/11/bits/invoke.h:111 ray-project#10 0x7ff2829fbadf in std::_Function_handler<void (ray::Status), ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Create<ray::rpc::NodeManagerService, ray::rpc::PinObjectIDsRequest, ray::rpc::PinObjectIDsReply>(std::weak_ptr<ray::rpc::RetryableGrpcClient>, std::unique_ptr<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply>, std::default_delete<grpc::ClientAsyncResponseReader<ray::rpc::PinObjectIDsReply> > > (ray::rpc::NodeManagerService::Stub::*)(grpc::ClientContext*, ray::rpc::PinObjectIDsRequest const&, grpc::CompletionQueue*), std::shared_ptr<ray::rpc::GrpcClient<ray::rpc::NodeManagerService> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, ray::rpc::PinObjectIDsRequest, std::function<void (ray::Status const&, ray::rpc::PinObjectIDsReply&&)>, long)::{lambda(ray::Status const&)ray-project#2}>::_M_invoke(std::_Any_data const&, ray::Status&&) /usr/include/c++/11/bits/std_function.h:290 ray-project#11 0x7ff2834657e9 in std::function<void (ray::Status)>::operator()(ray::Status) const /usr/include/c++/11/bits/std_function.h:590 ray-project#12 0x7ff2834657e9 in ray::rpc::RetryableGrpcClient::RetryableGrpcRequest::Fail(ray::Status const&) bazel-out/k8-dbg/bin/src/ray/rpc/_virtual_includes/retryable_grpc_client/ray/rpc/retryable_grpc_client.h:109 ray-project#13 0x7ff2834657e9 in operator() src/ray/rpc/retryable_grpc_client.cc:30 ray-project#14 0x7ff2834657e9 in __invoke_impl<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#15 0x7ff2834657e9 in __invoke_r<void, ray::rpc::RetryableGrpcClient::~RetryableGrpcClient()::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#16 0x7ff2834657e9 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 2025-11-14 16:15:05,608 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#17 0x7ff2834e2407 in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#18 0x7ff2834e2407 in EventTracker::RecordExecution(std::function<void ()> const&, std::shared_ptr<StatsHandle>) src/ray/common/event_stats.cc:112 ray-project#19 0x7ff2834bea54 in operator() src/ray/common/asio/instrumented_io_context.cc:110 ray-project#20 0x7ff2834bea54 in __invoke_impl<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#21 0x7ff2834bea54 in __invoke_r<void, instrumented_io_context::post(std::function<void()>, std::string, int64_t)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#22 0x7ff2834bea54 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#23 0x7ff28242fb5b in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#24 0x7ff28242fb5b in boost::asio::detail::binder0<std::function<void ()> >::operator()() external/boost/boost/asio/detail/bind_handler.hpp:60 ray-project#25 0x7ff28242fb5b in void boost::asio::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, ...) external/boost/boost/asio/handler_invoke_hook.hpp:88 ray-project#26 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, std::function<void ()>&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#27 0x7ff28242fb5b in void boost::asio::detail::asio_handler_invoke<boost::asio::detail::binder0<std::function<void ()> >, std::function<void ()> >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >*) external/boost/boost/asio/detail/bind_handler.hpp:111 ray-project#28 0x7ff28242fb5b in void boost_asio_handler_invoke_helpers::invoke<boost::asio::detail::binder0<std::function<void ()> >, boost::asio::detail::binder0<std::function<void ()> > >(boost::asio::detail::binder0<std::function<void ()> >&, boost::asio::detail::binder0<std::function<void ()> >&) external/boost/boost/asio/detail/handler_invoke_helpers.hpp:54 ray-project#29 0x7ff28242fb5b in boost::asio::detail::executor_op<boost::asio::detail::binder0<std::function<void ()> >, std::allocator<void>, boost::asio::detail::scheduler_operation>::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/executor_op.hpp:70 ray-project#30 0x7ff2838607d6 in boost::asio::detail::scheduler_operation::complete(void*, boost::system::error_code const&, unsigned long) external/boost/boost/asio/detail/scheduler_operation.hpp:40 ray-project#31 0x7ff2838607d6 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) external/boost/boost/asio/detail/impl/scheduler.ipp:492 ray-project#32 0x7ff283892d35 in boost::asio::detail::scheduler::run(boost::system::error_code&) external/boost/boost/asio/detail/impl/scheduler.ipp:210 ray-project#33 0x7ff2838981e0 in boost::asio::io_context::run() external/boost/boost/asio/impl/io_context.ipp:63 2025-11-14 16:15:05,742 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} ray-project#34 0x7ff281e9d0aa in operator() src/ray/core_worker/core_worker_process.cc:193 ray-project#35 0x7ff281e9d247 in run external/boost/boost/thread/detail/thread.hpp:120 ray-project#36 0x7ff282503c47 in thread_proxy external/boost/libs/thread/src/pthread/thread.cpp:179 ray-project#37 0x7ff28b013ac2 in start_thread nptl/pthread_create.c:442 ray-project#38 0x7ff28b0a58bf (/lib/x86_64-linux-gnu/libc.so.6+0x1268bf) 0x50c003fd3d30 is located 112 bytes inside of 120-byte region [0x50c003fd3cc0,0x50c003fd3d38) freed by thread T68 here: 2025-11-14 16:15:05,876 INFO streaming_executor_state.py:511 -- Running activate tasks is {'MapBatches(QwenInfer)': ['MapBatches(QwenInfer)-79153', 'MapBatches(QwenInfer)-80170', 'MapBatches(QwenInfer)-80225', 'MapBatches(QwenInfer)-80299', 'MapBatches(QwenInfer)-82624'], 'MapBatches(drop_columns)->Write': ['MapBatches(drop_columns)->Write-25244', 'MapBatches(drop_columns)->Write-34438', 'MapBatches(drop_columns)->Write-34439', 'MapBatches(drop_columns)->Write-34440', 'MapBatches(drop_columns)->Write-34441']} #0 0x7ff28b39924f in operator delete(void*, unsigned long) ../../../../src/libsanitizer/asan/asan_new_delete.cpp:172 ray-project#1 0x7ff281eceb5f in __gnu_cxx::new_allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >::deallocate(std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/ext/new_allocator.h:145 ray-project#2 0x7ff281eceb5f in std::allocator_traits<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::deallocate(std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> >&, std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>*, unsigned long) /usr/include/c++/11/bits/alloc_traits.h:496 ray-project#3 0x7ff281eceb5f in std::__allocated_ptr<std::allocator<std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2> > >::~__allocated_ptr() /usr/include/c++/11/bits/allocated_ptr.h:74 ray-project#4 0x7ff281eceb5f in std::_Sp_counted_ptr_inplace<ray::rpc::RayletClient, std::allocator<ray::rpc::RayletClient>, (__gnu_cxx::_Lock_policy)2>::_M_destroy() /usr/include/c++/11/bits/shared_ptr_base.h:538 ray-project#5 0x7ff282a73f0a in std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() /usr/include/c++/11/bits/shared_ptr_base.h:184 ray-project#6 0x7ff282a73f0a in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count() /usr/include/c++/11/bits/shared_ptr_base.h:705 ray-project#7 0x7ff282a73f0a in std::__shared_ptr<ray::RayletClientInterface, (__gnu_cxx::_Lock_policy)2>::~__shared_ptr() /usr/include/c++/11/bits/shared_ptr_base.h:1154 ray-project#8 0x7ff282a73f0a in std::shared_ptr<ray::RayletClientInterface>::~shared_ptr() /usr/include/c++/11/bits/shared_ptr.h:122 ray-project#9 0x7ff282a73f0a in std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >::~pair() /usr/include/c++/11/bits/stl_pair.h:211 ray-project#10 0x7ff282a73f0a in void __gnu_cxx::new_allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/ext/new_allocator.h:168 ray-project#11 0x7ff282a73f0a in void std::allocator_traits<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::destroy<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >&, std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> >*) /usr/include/c++/11/bits/alloc_traits.h:535 ray-project#12 0x7ff282a73f0a in void absl::lts_20230802::container_internal::map_slot_policy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/container_memory.h:421 ray-project#13 0x7ff282a73f0a in void absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/flat_hash_map.h:578 ray-project#14 0x7ff282a73f0a in void absl::lts_20230802::container_internal::common_policy_traits<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, void>::destroy<std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >(std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > >*, absl::lts_20230802::container_internal::map_slot_type<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >*) external/com_google_absl/absl/container/internal/common_policy_traits.h:50 ray-project#15 0x7ff282a73f0a in absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::erase(absl::lts_20230802::container_internal::raw_hash_set<absl::lts_20230802::container_internal::FlatHashMapPolicy<ray::NodeID, std::shared_ptr<ray::RayletClientInterface> >, absl::lts_20230802::hash_internal::Hash<ray::NodeID>, std::equal_to<ray::NodeID>, std::allocator<std::pair<ray::NodeID const, std::shared_ptr<ray::RayletClientInterface> > > >::iterator) external/com_google_absl/absl/container/internal/raw_hash_set.h:2183 ray-project#16 0x7ff282a73f0a in ray::rpc::RayletClientPool::Disconnect(ray::NodeID) src/ray/raylet_rpc_client/raylet_client_pool.cc:114 ray-project#17 0x7ff282a7aa61 in operator() src/ray/raylet_rpc_client/raylet_client_pool.cc:69 ray-project#18 0x7ff282a7ac66 in __invoke_impl<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:61 ray-project#19 0x7ff282a7ac66 in __invoke_r<void, ray::rpc::RayletClientPool::GetDefaultUnavailableTimeoutCallback(ray::gcs::GcsClient*, ray::rpc::RayletClientPool*, const ray::rpc::Address&)::<lambda()>&> /usr/include/c++/11/bits/invoke.h:111 ray-project#20 0x7ff282a7ac66 in _M_invoke /usr/include/c++/11/bits/std_function.h:290 ray-project#21 0x7ff28346a1ac in std::function<void ()>::operator()() const /usr/include/c++/11/bits/std_function.h:590 ray-project#22 0x7ff28346a1ac in ray::rpc::RetryableGrpcClient::CheckChannelStatus(bool) src/ray/rpc/retryable_grpc_client.cc:85 ray-project#23 0x7ff28346c06a in operator() src/ray/rpc/retryable_grpc_client.cc:45 ``` This is a **non-deterministic race condition** that occurs under the following sequence: 1. Worker A's pod is preempted → Worker A dies 2. Objects on Worker A are lost 3. Objects are found on Worker B → `PinObjectIDs` RPC is initiated 4. Worker B dies or becomes unavailable → `CheckChannelStatus` detects this → `Disconnect` is called 5. The `RayletClient` corresponding to Worker B on the driver is destroyed 6. RPC callback executes and accesses the already-freed `RayletClient` → use-after-free triggers crash Whether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers `PinObjectIDs`, making this race condition more likely to occur. ### Root Cause In `RayletClient::PinObjectIDs`, the RPC callback lambda directly captured the raw `this` pointer: ```cpp auto rpc_callback = [this, callback = std::move(callback)](...) { pins_in_flight_--; // Accessing member via 'this' pointer ... }; ``` If the `RayletClient` object is destroyed before the async RPC callback executes, the callback will access freed memory through the dangling `this` pointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size". ## Solution The fix ensures that the `RayletClient` object remains alive during the asynchronous callback execution by: 1. **Inheriting from `std::enable_shared_from_this<RayletClient>`**: The class already inherits from this base class (line 43 in `raylet_client.h`), which enables safe shared pointer management. 2. **Capturing `shared_from_this()` in the lambda**: Instead of capturing the raw `this` pointer, the callback now captures a `shared_ptr` to the object. The `shared_from_this()` is called before incrementing `pins_in_flight_` to ensure proper lifetime management: ```cpp // Capture shared_from_this() before incrementing to ensure object lifetime // is extended for the async callback, preventing use-after-free. auto self = shared_from_this(); pins_in_flight_++; auto rpc_callback = [self, callback = std::move(callback)]( Status status, rpc::PinObjectIDsReply &&reply) { self->pins_in_flight_--; callback(status, std::move(reply)); }; ``` This ensures that the `RayletClient` object's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that if `shared_from_this()` were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state. ## Code Changes - **File**: `src/ray/raylet_rpc_client/raylet_client.cc` - **Method**: `RayletClient::PinObjectIDs` - **Change**: Replace `this` capture with `shared_from_this()` capture in the RPC callback lambda Signed-off-by: dragongu <andrewgu@vip.qq.com> Co-authored-by: gulonglong <gulonglong@stepfun.com> Signed-off-by: peterxcli <peterxcli@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR fixes a heap corruption bug that causes the driver to crash with SIGABRT. The issue is caused by a use-after-free when the
RayletClientobject is destroyed while an asynchronous RPC callback is still pending.Problem Description
Scenario
A Ray Data job (Ray 2.50.0) with pipeline
read_parquet -> filter -> map_batches -> writerunning for 4+ hours, where workers use elastic resources with low job priority causing frequent worker deaths due to pod preemption, crashes the driver with SIGABRT:Trigger Conditions
After reproducing with an ASan image, Asan reveals the actual use-after-free at:
This is a non-deterministic race condition that occurs under the following sequence:
PinObjectIDsRPC is initiatedCheckChannelStatusdetects this →Disconnectis calledRayletClientcorresponding to Worker B on the driver is destroyedRayletClient→ use-after-free triggers crashWhether the use-after-free occurs depends on the relative timing of steps 5 and 6. In scenarios with frequent pod preemptions, object recovery frequently triggers
PinObjectIDs, making this race condition more likely to occur.Root Cause
In
RayletClient::PinObjectIDs, the RPC callback lambda directly captured the rawthispointer:If the
RayletClientobject is destroyed before the async RPC callback executes, the callback will access freed memory through the danglingthispointer, leading to heap corruption and SIGABRT with the error message "corrupted size vs. prev_size".Solution
The fix ensures that the
RayletClientobject remains alive during the asynchronous callback execution by:Inheriting from
std::enable_shared_from_this<RayletClient>: The class already inherits from this base class (line 43 inraylet_client.h), which enables safe shared pointer management.Capturing
shared_from_this()in the lambda: Instead of capturing the rawthispointer, the callback now captures ashared_ptrto the object. Theshared_from_this()is called before incrementingpins_in_flight_to ensure proper lifetime management:This ensures that the
RayletClientobject's lifetime is extended until the callback completes, preventing the use-after-free bug. By capturing the shared pointer before incrementing the counter, we also ensure that ifshared_from_this()were to fail (though it shouldn't in normal usage), we don't leave the counter in an inconsistent state.Code Changes
src/ray/raylet_rpc_client/raylet_client.ccRayletClient::PinObjectIDsthiscapture withshared_from_this()capture in the RPC callback lambda