Skip to content

Conversation

@dependabot
Copy link

@dependabot dependabot bot commented on behalf of github May 19, 2021

Bumps tensorflow from 2.4.1 to 2.5.0.

Release notes

Sourced from tensorflow's releases.

TensorFlow 2.5.0

Release 2.5.0

Major Features and Improvements

  • Support for Python3.9 has been added.
  • tf.data:
    • tf.data service now supports strict round-robin reads, which is useful for synchronous training workloads where example sizes vary. With strict round robin reads, users can guarantee that consumers get similar-sized examples in the same step.
    • tf.data service now supports optional compression. Previously data would always be compressed, but now you can disable compression by passing compression=None to tf.data.experimental.service.distribute(...).
    • tf.data.Dataset.batch() now supports num_parallel_calls and deterministic arguments. num_parallel_calls is used to indicate that multiple input batches should be computed in parallel. With num_parallel_calls set, deterministic is used to indicate that outputs can be obtained in the non-deterministic order.
    • Options returned by tf.data.Dataset.options() are no longer mutable.
    • tf.data input pipelines can now be executed in debug mode, which disables any asynchrony, parallelism, or non-determinism and forces Python execution (as opposed to trace-compiled graph execution) of user-defined functions passed into transformations such as map. The debug mode can be enabled through tf.data.experimental.enable_debug_mode().
  • tf.lite
    • Enabled the new MLIR-based quantization backend by default
      • The new backend is used for 8 bits full integer post-training quantization
      • The new backend removes the redundant rescales and fixes some bugs (shared weight/bias, extremely small scales, etc)
      • Set experimental_new_quantizer in tf.lite.TFLiteConverter to False to disable this change
  • tf.keras
    • tf.keras.metrics.AUC now support logit predictions.
    • Enabled a new supported input type in Model.fit, tf.keras.utils.experimental.DatasetCreator, which takes a callable, dataset_fn. DatasetCreator is intended to work across all tf.distribute strategies, and is the only input type supported for Parameter Server strategy.
  • tf.distribute
    • tf.distribute.experimental.ParameterServerStrategy now supports training with Keras Model.fit when used with DatasetCreator.
    • Creating tf.random.Generator under tf.distribute.Strategy scopes is now allowed (except for tf.distribute.experimental.CentralStorageStrategy and tf.distribute.experimental.ParameterServerStrategy). Different replicas will get different random-number streams.
  • TPU embedding support
    • Added profile_data_directory to EmbeddingConfigSpec in _tpu_estimator_embedding.py. This allows embedding lookup statistics gathered at runtime to be used in embedding layer partitioning decisions.
  • PluggableDevice
  • oneAPI Deep Neural Network Library (oneDNN) CPU performance optimizations from Intel-optimized TensorFlow are now available in the official x86-64 Linux and Windows builds.
    • They are off by default. Enable them by setting the environment variable TF_ENABLE_ONEDNN_OPTS=1.
    • We do not recommend using them in GPU systems, as they have not been sufficiently tested with GPUs yet.
  • TensorFlow pip packages are now built with CUDA11.2 and cuDNN 8.1.0

Breaking Changes

  • The TF_CPP_MIN_VLOG_LEVEL environment variable has been renamed to to TF_CPP_MAX_VLOG_LEVEL which correctly describes its effect.

Bug Fixes and Other Changes

  • tf.keras:
    • Preprocessing layers API consistency changes:
      • StringLookup added output_mode, sparse, and pad_to_max_tokens arguments with same semantics as TextVectorization.
      • IntegerLookup added output_mode, sparse, and pad_to_max_tokens arguments with same semantics as TextVectorization. Renamed max_values, oov_value and mask_value to max_tokens, oov_token and mask_token to align with StringLookup and TextVectorization.
      • TextVectorization default for pad_to_max_tokens switched to False.
      • CategoryEncoding no longer supports adapt, IntegerLookup now supports equivalent functionality. max_tokens argument renamed to num_tokens.
      • Discretization added num_bins argument for learning bins boundaries through calling adapt on a dataset. Renamed bins argument to bin_boundaries for specifying bins without adapt.
    • Improvements to model saving/loading:
      • model.load_weights now accepts paths to saved models.

... (truncated)

Changelog

Sourced from tensorflow's changelog.

Release 2.5.0

Breaking Changes

  • The TF_CPP_MIN_VLOG_LEVEL environment variable has been renamed to to TF_CPP_MAX_VLOG_LEVEL which correctly describes its effect.

Known Caveats

Major Features and Improvements

  • TPU embedding support

    • Added profile_data_directory to EmbeddingConfigSpec in _tpu_estimator_embedding.py. This allows embedding lookup statistics gathered at runtime to be used in embedding layer partitioning decisions.
  • tf.keras.metrics.AUC now support logit predictions.

  • Creating tf.random.Generator under tf.distribute.Strategy scopes is now allowed (except for tf.distribute.experimental.CentralStorageStrategy and tf.distribute.experimental.ParameterServerStrategy). Different replicas will get different random-number streams.

  • tf.data:

    • tf.data service now supports strict round-robin reads, which is useful for synchronous training workloads where example sizes vary. With strict round robin reads, users can guarantee that consumers get similar-sized examples in the same step.
    • tf.data service now supports optional compression. Previously data would always be compressed, but now you can disable compression by passing compression=None to tf.data.experimental.service.distribute(...).
    • tf.data.Dataset.batch() now supports num_parallel_calls and deterministic arguments. num_parallel_calls is used to indicate that multiple input batches should be computed in parallel. With num_parallel_calls set, deterministic is used to indicate that outputs can be obtained in the non-deterministic order.
    • Options returned by tf.data.Dataset.options() are no longer mutable.
    • tf.data input pipelines can now be executed in debug mode, which disables any asynchrony, parallelism, or non-determinism and forces Python execution (as opposed to trace-compiled graph execution) of user-defined functions passed into transformations such as map. The debug mode can be enabled through tf.data.experimental.enable_debug_mode().
  • tf.lite

    • Enabled the new MLIR-based quantization backend by default
      • The new backend is used for 8 bits full integer post-training quantization
      • The new backend removes the redundant rescales and fixes some bugs (shared weight/bias, extremely small scales, etc)

... (truncated)

Commits
  • a4dfb8d Merge pull request #49124 from tensorflow/mm-cherrypick-tf-data-segfault-fix-...
  • 2107b1d Merge pull request #49116 from tensorflow-jenkins/version-numbers-2.5.0-17609
  • 16b8139 Update snapshot_dataset_op.cc
  • 86a0d86 Merge pull request #49126 from geetachavan1/cherrypicks_X9ZNY
  • 9436ae6 Merge pull request #49128 from geetachavan1/cherrypicks_D73J5
  • 6b2bf99 Validate that a and b are proper sparse tensors
  • c03ad1a Ensure validation sticks in banded_triangular_solve_op
  • 12a6ead Merge pull request #49120 from geetachavan1/cherrypicks_KJ5M9
  • b67f5b8 Merge pull request #49118 from geetachavan1/cherrypicks_BIDTR
  • a13c0ad [tf.data][cherrypick] Fix snapshot segfault when using repeat and prefecth
  • Additional commits viewable in compare view

Dependabot compatibility score

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.


Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

  • @dependabot rebase will rebase this PR
  • @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
  • @dependabot merge will merge this PR after your CI passes on it
  • @dependabot squash and merge will squash and merge this PR after your CI passes on it
  • @dependabot cancel merge will cancel a previously requested merge and block automerging
  • @dependabot reopen will reopen this PR if it is closed
  • @dependabot close will close this PR and stop Dependabot recreating it. You can achieve the same result by closing it manually
  • @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
  • @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
  • @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

@dependabot dependabot bot added the dependencies Pull requests that update a dependency file label May 19, 2021
@dependabot dependabot bot force-pushed the dependabot/pip/python/requirements/rllib/tensorflow-2.5.0 branch from 2e7f286 to 6352c26 Compare May 24, 2021 19:33
@dependabot @github
Copy link
Author

dependabot bot commented on behalf of github Aug 14, 2021

Superseded by #48.

@dependabot dependabot bot closed this Aug 14, 2021
@dependabot dependabot bot deleted the dependabot/pip/python/requirements/rllib/tensorflow-2.5.0 branch August 14, 2021 07:06
edoakes pushed a commit 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 <[email protected]>
Co-authored-by: gulonglong <[email protected]>
edoakes pushed a commit 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 <[email protected]>
Co-authored-by: gulonglong <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

dependencies Pull requests that update a dependency file

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant