diff --git a/include/grpcpp/impl/codegen/client_callback_impl.h b/include/grpcpp/impl/codegen/client_callback_impl.h
index 8e683743e06..86a46d8c870 100644
--- a/include/grpcpp/impl/codegen/client_callback_impl.h
+++ b/include/grpcpp/impl/codegen/client_callback_impl.h
@@ -461,76 +461,51 @@ class ClientCallbackReaderWriterImpl
     // 1. Send initial metadata (unless corked) + recv initial metadata
     // 2. Any read backlog
     // 3. Any write backlog
-    // 4. Recv trailing metadata, on_completion callback
-    started_ = true;
-
-    start_tag_.Set(call_.call(),
-                   [this](bool ok) {
-                     reactor_->OnReadInitialMetadataDone(ok);
-                     MaybeFinish();
-                   },
-                   &start_ops_, /*can_inline=*/false);
+    // 4. Recv trailing metadata (unless corked)
     if (!start_corked_) {
       start_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
                                      context_->initial_metadata_flags());
     }
-    start_ops_.RecvInitialMetadata(context_);
-    start_ops_.set_core_cq_tag(&start_tag_);
-    call_.PerformOps(&start_ops_);
-
-    // Also set up the read and write tags so that they don't have to be set up
-    // each time
-    write_tag_.Set(call_.call(),
-                   [this](bool ok) {
-                     reactor_->OnWriteDone(ok);
-                     MaybeFinish();
-                   },
-                   &write_ops_, /*can_inline=*/false);
-    write_ops_.set_core_cq_tag(&write_tag_);
-
-    read_tag_.Set(call_.call(),
-                  [this](bool ok) {
-                    reactor_->OnReadDone(ok);
-                    MaybeFinish();
-                  },
-                  &read_ops_, /*can_inline=*/false);
-    read_ops_.set_core_cq_tag(&read_tag_);
-    if (read_ops_at_start_) {
-      call_.PerformOps(&read_ops_);
-    }
 
-    if (write_ops_at_start_) {
-      call_.PerformOps(&write_ops_);
-    }
+    call_.PerformOps(&start_ops_);
 
-    if (writes_done_ops_at_start_) {
-      call_.PerformOps(&writes_done_ops_);
+    {
+      grpc::internal::MutexLock lock(&start_mu_);
+
+      if (backlog_.read_ops) {
+        call_.PerformOps(&read_ops_);
+      }
+      if (backlog_.write_ops) {
+        call_.PerformOps(&write_ops_);
+      }
+      if (backlog_.writes_done_ops) {
+        call_.PerformOps(&writes_done_ops_);
+      }
+      call_.PerformOps(&finish_ops_);
+      // The last thing in this critical section is to set started_ so that it
+      // can be used lock-free as well.
+      started_.store(true, std::memory_order_release);
     }
-
-    finish_tag_.Set(call_.call(), [this](bool /*ok*/) { MaybeFinish(); },
-                    &finish_ops_, /*can_inline=*/false);
-    finish_ops_.ClientRecvStatus(context_, &finish_status_);
-    finish_ops_.set_core_cq_tag(&finish_tag_);
-    call_.PerformOps(&finish_ops_);
+    // MaybeFinish outside the lock to make sure that destruction of this object
+    // doesn't take place while holding the lock (which would cause the lock to
+    // be released after destruction)
+    this->MaybeFinish();
   }
 
   void Read(Response* msg) override {
     read_ops_.RecvMessage(msg);
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&read_ops_);
-    } else {
-      read_ops_at_start_ = true;
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.read_ops = true;
+        return;
+      }
     }
+    call_.PerformOps(&read_ops_);
   }
 
   void Write(const Request* msg, ::grpc::WriteOptions options) override {
-    if (start_corked_) {
-      write_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
-                                     context_->initial_metadata_flags());
-      start_corked_ = false;
-    }
-
     if (options.is_last_message()) {
       options.set_buffer_hint();
       write_ops_.ClientSendClose();
@@ -538,18 +513,22 @@ class ClientCallbackReaderWriterImpl
     // TODO(vjpai): don't assert
     GPR_CODEGEN_ASSERT(write_ops_.SendMessagePtr(msg, options).ok());
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&write_ops_);
-    } else {
-      write_ops_at_start_ = true;
+    if (GPR_UNLIKELY(corked_write_needed_)) {
+      write_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
+                                     context_->initial_metadata_flags());
+      corked_write_needed_ = false;
+    }
+
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.write_ops = true;
+        return;
+      }
     }
+    call_.PerformOps(&write_ops_);
   }
   void WritesDone() override {
-    if (start_corked_) {
-      writes_done_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
-                                           context_->initial_metadata_flags());
-      start_corked_ = false;
-    }
     writes_done_ops_.ClientSendClose();
     writes_done_tag_.Set(call_.call(),
                          [this](bool ok) {
@@ -559,11 +538,19 @@ class ClientCallbackReaderWriterImpl
                          &writes_done_ops_, /*can_inline=*/false);
     writes_done_ops_.set_core_cq_tag(&writes_done_tag_);
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&writes_done_ops_);
-    } else {
-      writes_done_ops_at_start_ = true;
+    if (GPR_UNLIKELY(corked_write_needed_)) {
+      writes_done_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
+                                           context_->initial_metadata_flags());
+      corked_write_needed_ = false;
     }
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.writes_done_ops = true;
+        return;
+      }
+    }
+    call_.PerformOps(&writes_done_ops_);
   }
 
   void AddHold(int holds) override {
@@ -580,8 +567,42 @@ class ClientCallbackReaderWriterImpl
       : context_(context),
         call_(call),
         reactor_(reactor),
-        start_corked_(context_->initial_metadata_corked_) {
+        start_corked_(context_->initial_metadata_corked_),
+        corked_write_needed_(start_corked_) {
     this->BindReactor(reactor);
+
+    // Set up the unchanging parts of the start, read, and write tags and ops.
+    start_tag_.Set(call_.call(),
+                   [this](bool ok) {
+                     reactor_->OnReadInitialMetadataDone(ok);
+                     MaybeFinish();
+                   },
+                   &start_ops_, /*can_inline=*/false);
+    start_ops_.RecvInitialMetadata(context_);
+    start_ops_.set_core_cq_tag(&start_tag_);
+
+    write_tag_.Set(call_.call(),
+                   [this](bool ok) {
+                     reactor_->OnWriteDone(ok);
+                     MaybeFinish();
+                   },
+                   &write_ops_, /*can_inline=*/false);
+    write_ops_.set_core_cq_tag(&write_tag_);
+
+    read_tag_.Set(call_.call(),
+                  [this](bool ok) {
+                    reactor_->OnReadDone(ok);
+                    MaybeFinish();
+                  },
+                  &read_ops_, /*can_inline=*/false);
+    read_ops_.set_core_cq_tag(&read_tag_);
+
+    // Also set up the Finish tag and op set.
+    finish_tag_.Set(call_.call(), [this](bool /*ok*/) { MaybeFinish(); },
+                    &finish_ops_,
+                    /*can_inline=*/false);
+    finish_ops_.ClientRecvStatus(context_, &finish_status_);
+    finish_ops_.set_core_cq_tag(&finish_tag_);
   }
 
   ::grpc_impl::ClientContext* const context_;
@@ -592,7 +613,9 @@ class ClientCallbackReaderWriterImpl
                             grpc::internal::CallOpRecvInitialMetadata>
       start_ops_;
   grpc::internal::CallbackWithSuccessTag start_tag_;
-  bool start_corked_;
+  const bool start_corked_;
+  bool corked_write_needed_;  // no lock needed since only accessed in
+                              // Write/WritesDone which cannot be concurrent
 
   grpc::internal::CallOpSet<grpc::internal::CallOpClientRecvStatus> finish_ops_;
   grpc::internal::CallbackWithSuccessTag finish_tag_;
@@ -603,22 +626,27 @@ class ClientCallbackReaderWriterImpl
                             grpc::internal::CallOpClientSendClose>
       write_ops_;
   grpc::internal::CallbackWithSuccessTag write_tag_;
-  bool write_ops_at_start_{false};
 
   grpc::internal::CallOpSet<grpc::internal::CallOpSendInitialMetadata,
                             grpc::internal::CallOpClientSendClose>
       writes_done_ops_;
   grpc::internal::CallbackWithSuccessTag writes_done_tag_;
-  bool writes_done_ops_at_start_{false};
 
   grpc::internal::CallOpSet<grpc::internal::CallOpRecvMessage<Response>>
       read_ops_;
   grpc::internal::CallbackWithSuccessTag read_tag_;
-  bool read_ops_at_start_{false};
 
-  // Minimum of 2 callbacks to pre-register for start and finish
-  std::atomic<intptr_t> callbacks_outstanding_{2};
-  bool started_{false};
+  struct StartCallBacklog {
+    bool write_ops = false;
+    bool writes_done_ops = false;
+    bool read_ops = false;
+  };
+  StartCallBacklog backlog_ /* GUARDED_BY(start_mu_) */;
+
+  // Minimum of 3 callbacks to pre-register for start ops, StartCall, and finish
+  std::atomic<intptr_t> callbacks_outstanding_{3};
+  std::atomic_bool started_{false};
+  grpc::internal::Mutex start_mu_;
 };
 
 template <class Request, class Response>
@@ -670,8 +698,7 @@ class ClientCallbackReaderImpl : public ClientCallbackReader<Response> {
     // This call initiates two batches, plus any backlog, each with a callback
     // 1. Send initial metadata (unless corked) + recv initial metadata
     // 2. Any backlog
-    // 3. Recv trailing metadata, on_completion callback
-    started_ = true;
+    // 3. Recv trailing metadata
 
     start_tag_.Set(call_.call(),
                    [this](bool ok) {
@@ -693,8 +720,13 @@ class ClientCallbackReaderImpl : public ClientCallbackReader<Response> {
                   },
                   &read_ops_, /*can_inline=*/false);
     read_ops_.set_core_cq_tag(&read_tag_);
-    if (read_ops_at_start_) {
-      call_.PerformOps(&read_ops_);
+
+    {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (backlog_.read_ops) {
+        call_.PerformOps(&read_ops_);
+      }
+      started_.store(true, std::memory_order_release);
     }
 
     finish_tag_.Set(call_.call(), [this](bool /*ok*/) { MaybeFinish(); },
@@ -707,11 +739,14 @@ class ClientCallbackReaderImpl : public ClientCallbackReader<Response> {
   void Read(Response* msg) override {
     read_ops_.RecvMessage(msg);
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&read_ops_);
-    } else {
-      read_ops_at_start_ = true;
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.read_ops = true;
+        return;
+      }
     }
+    call_.PerformOps(&read_ops_);
   }
 
   void AddHold(int holds) override {
@@ -752,11 +787,16 @@ class ClientCallbackReaderImpl : public ClientCallbackReader<Response> {
   grpc::internal::CallOpSet<grpc::internal::CallOpRecvMessage<Response>>
       read_ops_;
   grpc::internal::CallbackWithSuccessTag read_tag_;
-  bool read_ops_at_start_{false};
+
+  struct StartCallBacklog {
+    bool read_ops = false;
+  };
+  StartCallBacklog backlog_ /* GUARDED_BY(start_mu_) */;
 
   // Minimum of 2 callbacks to pre-register for start and finish
   std::atomic<intptr_t> callbacks_outstanding_{2};
-  bool started_{false};
+  std::atomic_bool started_{false};
+  grpc::internal::Mutex start_mu_;
 };
 
 template <class Response>
@@ -809,74 +849,60 @@ class ClientCallbackWriterImpl : public ClientCallbackWriter<Request> {
     // This call initiates two batches, plus any backlog, each with a callback
     // 1. Send initial metadata (unless corked) + recv initial metadata
     // 2. Any backlog
-    // 3. Recv trailing metadata, on_completion callback
-    started_ = true;
+    // 3. Recv trailing metadata
 
-    start_tag_.Set(call_.call(),
-                   [this](bool ok) {
-                     reactor_->OnReadInitialMetadataDone(ok);
-                     MaybeFinish();
-                   },
-                   &start_ops_, /*can_inline=*/false);
     if (!start_corked_) {
       start_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
                                      context_->initial_metadata_flags());
     }
-    start_ops_.RecvInitialMetadata(context_);
-    start_ops_.set_core_cq_tag(&start_tag_);
     call_.PerformOps(&start_ops_);
 
-    // Also set up the read and write tags so that they don't have to be set up
-    // each time
-    write_tag_.Set(call_.call(),
-                   [this](bool ok) {
-                     reactor_->OnWriteDone(ok);
-                     MaybeFinish();
-                   },
-                   &write_ops_, /*can_inline=*/false);
-    write_ops_.set_core_cq_tag(&write_tag_);
-
-    if (write_ops_at_start_) {
-      call_.PerformOps(&write_ops_);
+    {
+      grpc::internal::MutexLock lock(&start_mu_);
+
+      if (backlog_.write_ops) {
+        call_.PerformOps(&write_ops_);
+      }
+      if (backlog_.writes_done_ops) {
+        call_.PerformOps(&writes_done_ops_);
+      }
+      call_.PerformOps(&finish_ops_);
+      // The last thing in this critical section is to set started_ so that it
+      // can be used lock-free as well.
+      started_.store(true, std::memory_order_release);
     }
-
-    if (writes_done_ops_at_start_) {
-      call_.PerformOps(&writes_done_ops_);
-    }
-
-    finish_tag_.Set(call_.call(), [this](bool /*ok*/) { MaybeFinish(); },
-                    &finish_ops_, /*can_inline=*/false);
-    finish_ops_.ClientRecvStatus(context_, &finish_status_);
-    finish_ops_.set_core_cq_tag(&finish_tag_);
-    call_.PerformOps(&finish_ops_);
+    // MaybeFinish outside the lock to make sure that destruction of this object
+    // doesn't take place while holding the lock (which would cause the lock to
+    // be released after destruction)
+    this->MaybeFinish();
   }
 
   void Write(const Request* msg, ::grpc::WriteOptions options) override {
-    if (start_corked_) {
-      write_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
-                                     context_->initial_metadata_flags());
-      start_corked_ = false;
-    }
-
-    if (options.is_last_message()) {
+    if (GPR_UNLIKELY(options.is_last_message())) {
       options.set_buffer_hint();
       write_ops_.ClientSendClose();
     }
     // TODO(vjpai): don't assert
     GPR_CODEGEN_ASSERT(write_ops_.SendMessagePtr(msg, options).ok());
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&write_ops_);
-    } else {
-      write_ops_at_start_ = true;
+
+    if (GPR_UNLIKELY(corked_write_needed_)) {
+      write_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
+                                     context_->initial_metadata_flags());
+      corked_write_needed_ = false;
     }
+
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.write_ops = true;
+        return;
+      }
+    }
+    call_.PerformOps(&write_ops_);
   }
+
   void WritesDone() override {
-    if (start_corked_) {
-      writes_done_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
-                                           context_->initial_metadata_flags());
-      start_corked_ = false;
-    }
     writes_done_ops_.ClientSendClose();
     writes_done_tag_.Set(call_.call(),
                          [this](bool ok) {
@@ -886,11 +912,21 @@ class ClientCallbackWriterImpl : public ClientCallbackWriter<Request> {
                          &writes_done_ops_, /*can_inline=*/false);
     writes_done_ops_.set_core_cq_tag(&writes_done_tag_);
     callbacks_outstanding_.fetch_add(1, std::memory_order_relaxed);
-    if (started_) {
-      call_.PerformOps(&writes_done_ops_);
-    } else {
-      writes_done_ops_at_start_ = true;
+
+    if (GPR_UNLIKELY(corked_write_needed_)) {
+      writes_done_ops_.SendInitialMetadata(&context_->send_initial_metadata_,
+                                           context_->initial_metadata_flags());
+      corked_write_needed_ = false;
+    }
+
+    if (GPR_UNLIKELY(!started_.load(std::memory_order_acquire))) {
+      grpc::internal::MutexLock lock(&start_mu_);
+      if (GPR_LIKELY(!started_.load(std::memory_order_relaxed))) {
+        backlog_.writes_done_ops = true;
+        return;
+      }
     }
+    call_.PerformOps(&writes_done_ops_);
   }
 
   void AddHold(int holds) override {
@@ -909,10 +945,36 @@ class ClientCallbackWriterImpl : public ClientCallbackWriter<Request> {
       : context_(context),
         call_(call),
         reactor_(reactor),
-        start_corked_(context_->initial_metadata_corked_) {
+        start_corked_(context_->initial_metadata_corked_),
+        corked_write_needed_(start_corked_) {
     this->BindReactor(reactor);
+
+    // Set up the unchanging parts of the start and write tags and ops.
+    start_tag_.Set(call_.call(),
+                   [this](bool ok) {
+                     reactor_->OnReadInitialMetadataDone(ok);
+                     MaybeFinish();
+                   },
+                   &start_ops_, /*can_inline=*/false);
+    start_ops_.RecvInitialMetadata(context_);
+    start_ops_.set_core_cq_tag(&start_tag_);
+
+    write_tag_.Set(call_.call(),
+                   [this](bool ok) {
+                     reactor_->OnWriteDone(ok);
+                     MaybeFinish();
+                   },
+                   &write_ops_, /*can_inline=*/false);
+    write_ops_.set_core_cq_tag(&write_tag_);
+
+    // Also set up the Finish tag and op set.
     finish_ops_.RecvMessage(response);
     finish_ops_.AllowNoMessage();
+    finish_tag_.Set(call_.call(), [this](bool /*ok*/) { MaybeFinish(); },
+                    &finish_ops_,
+                    /*can_inline=*/false);
+    finish_ops_.ClientRecvStatus(context_, &finish_status_);
+    finish_ops_.set_core_cq_tag(&finish_tag_);
   }
 
   ::grpc_impl::ClientContext* const context_;
@@ -923,7 +985,9 @@ class ClientCallbackWriterImpl : public ClientCallbackWriter<Request> {
                             grpc::internal::CallOpRecvInitialMetadata>
       start_ops_;
   grpc::internal::CallbackWithSuccessTag start_tag_;
-  bool start_corked_;
+  const bool start_corked_;
+  bool corked_write_needed_;  // no lock needed since only accessed in
+                              // Write/WritesDone which cannot be concurrent
 
   grpc::internal::CallOpSet<grpc::internal::CallOpGenericRecvMessage,
                             grpc::internal::CallOpClientRecvStatus>
@@ -936,17 +1000,22 @@ class ClientCallbackWriterImpl : public ClientCallbackWriter<Request> {
                             grpc::internal::CallOpClientSendClose>
       write_ops_;
   grpc::internal::CallbackWithSuccessTag write_tag_;
-  bool write_ops_at_start_{false};
 
   grpc::internal::CallOpSet<grpc::internal::CallOpSendInitialMetadata,
                             grpc::internal::CallOpClientSendClose>
       writes_done_ops_;
   grpc::internal::CallbackWithSuccessTag writes_done_tag_;
-  bool writes_done_ops_at_start_{false};
 
-  // Minimum of 2 callbacks to pre-register for start and finish
-  std::atomic<intptr_t> callbacks_outstanding_{2};
-  bool started_{false};
+  struct StartCallBacklog {
+    bool write_ops = false;
+    bool writes_done_ops = false;
+  };
+  StartCallBacklog backlog_ /* GUARDED_BY(start_mu_) */;
+
+  // Minimum of 3 callbacks to pre-register for start ops, StartCall, and finish
+  std::atomic<intptr_t> callbacks_outstanding_{3};
+  std::atomic_bool started_{false};
+  grpc::internal::Mutex start_mu_;
 };
 
 template <class Request>
@@ -985,7 +1054,6 @@ class ClientCallbackUnaryImpl final : public ClientCallbackUnary {
     // This call initiates two batches, each with a callback
     // 1. Send initial metadata + write + writes done + recv initial metadata
     // 2. Read message, recv trailing metadata
-    started_ = true;
 
     start_tag_.Set(call_.call(),
                    [this](bool ok) {
@@ -1053,7 +1121,6 @@ class ClientCallbackUnaryImpl final : public ClientCallbackUnary {
 
   // This call will have 2 callbacks: start and finish
   std::atomic<intptr_t> callbacks_outstanding_{2};
-  bool started_{false};
 };
 
 class ClientCallbackUnaryFactory {
diff --git a/test/cpp/end2end/client_callback_end2end_test.cc b/test/cpp/end2end/client_callback_end2end_test.cc
index 05aac336789..cd3c14881bd 100644
--- a/test/cpp/end2end/client_callback_end2end_test.cc
+++ b/test/cpp/end2end/client_callback_end2end_test.cc
@@ -16,12 +16,6 @@
  *
  */
 
-#include <algorithm>
-#include <functional>
-#include <mutex>
-#include <sstream>
-#include <thread>
-
 #include <grpcpp/channel.h>
 #include <grpcpp/client_context.h>
 #include <grpcpp/create_channel.h>
@@ -31,6 +25,14 @@
 #include <grpcpp/server_builder.h>
 #include <grpcpp/server_context.h>
 #include <grpcpp/support/client_callback.h>
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <condition_variable>
+#include <functional>
+#include <mutex>
+#include <sstream>
+#include <thread>
 
 #include "src/core/lib/gpr/env.h"
 #include "src/core/lib/iomgr/iomgr.h"
@@ -43,8 +45,6 @@
 #include "test/cpp/util/string_ref_helper.h"
 #include "test/cpp/util/test_credentials_provider.h"
 
-#include <gtest/gtest.h>
-
 // MAYBE_SKIP_TEST is a macro to determine if this particular test configuration
 // should be skipped based on a decision made at SetUp time. In particular, any
 // callback tests can only be run if the iomgr can run in the background or if
@@ -1114,7 +1114,8 @@ class BidiClient
  public:
   BidiClient(grpc::testing::EchoTestService::Stub* stub,
              ServerTryCancelRequestPhase server_try_cancel,
-             int num_msgs_to_send, ClientCancelInfo client_cancel = {})
+             int num_msgs_to_send, bool cork_metadata, bool first_write_async,
+             ClientCancelInfo client_cancel = {})
       : server_try_cancel_(server_try_cancel),
         msgs_to_send_{num_msgs_to_send},
         client_cancel_{client_cancel} {
@@ -1124,8 +1125,9 @@ class BidiClient
                            grpc::to_string(server_try_cancel));
     }
     request_.set_message("Hello fren ");
+    context_.set_initial_metadata_corked(cork_metadata);
     stub->experimental_async()->BidiStream(&context_, this);
-    MaybeWrite();
+    MaybeAsyncWrite(first_write_async);
     StartRead(&response_);
     StartCall();
   }
@@ -1146,6 +1148,10 @@ class BidiClient
     }
   }
   void OnWriteDone(bool ok) override {
+    if (async_write_thread_.joinable()) {
+      async_write_thread_.join();
+      RemoveHold();
+    }
     if (server_try_cancel_ == DO_NOT_CANCEL) {
       EXPECT_TRUE(ok);
     } else if (!ok) {
@@ -1210,6 +1216,26 @@ class BidiClient
   }
 
  private:
+  void MaybeAsyncWrite(bool first_write_async) {
+    if (first_write_async) {
+      // Make sure that we have a write to issue.
+      // TODO(vjpai): Make this work with 0 writes case as well.
+      assert(msgs_to_send_ >= 1);
+
+      AddHold();
+      async_write_thread_ = std::thread([this] {
+        std::unique_lock<std::mutex> lock(async_write_thread_mu_);
+        async_write_thread_cv_.wait(
+            lock, [this] { return async_write_thread_start_; });
+        MaybeWrite();
+      });
+      std::lock_guard<std::mutex> lock(async_write_thread_mu_);
+      async_write_thread_start_ = true;
+      async_write_thread_cv_.notify_one();
+      return;
+    }
+    MaybeWrite();
+  }
   void MaybeWrite() {
     if (client_cancel_.cancel &&
         writes_complete_ == client_cancel_.ops_before_cancel) {
@@ -1231,13 +1257,57 @@ class BidiClient
   std::mutex mu_;
   std::condition_variable cv_;
   bool done_ = false;
+  std::thread async_write_thread_;
+  bool async_write_thread_start_ = false;
+  std::mutex async_write_thread_mu_;
+  std::condition_variable async_write_thread_cv_;
 };
 
 TEST_P(ClientCallbackEnd2endTest, BidiStream) {
   MAYBE_SKIP_TEST;
   ResetStub();
-  BidiClient test{stub_.get(), DO_NOT_CANCEL,
-                  kServerDefaultResponseStreamsToSend};
+  BidiClient test(stub_.get(), DO_NOT_CANCEL,
+                  kServerDefaultResponseStreamsToSend,
+                  /*cork_metadata=*/false, /*first_write_async=*/false);
+  test.Await();
+  // Make sure that the server interceptors were not notified of a cancel
+  if (GetParam().use_interceptors) {
+    EXPECT_EQ(0, DummyInterceptor::GetNumTimesCancel());
+  }
+}
+
+TEST_P(ClientCallbackEnd2endTest, BidiStreamFirstWriteAsync) {
+  MAYBE_SKIP_TEST;
+  ResetStub();
+  BidiClient test(stub_.get(), DO_NOT_CANCEL,
+                  kServerDefaultResponseStreamsToSend,
+                  /*cork_metadata=*/false, /*first_write_async=*/true);
+  test.Await();
+  // Make sure that the server interceptors were not notified of a cancel
+  if (GetParam().use_interceptors) {
+    EXPECT_EQ(0, DummyInterceptor::GetNumTimesCancel());
+  }
+}
+
+TEST_P(ClientCallbackEnd2endTest, BidiStreamCorked) {
+  MAYBE_SKIP_TEST;
+  ResetStub();
+  BidiClient test(stub_.get(), DO_NOT_CANCEL,
+                  kServerDefaultResponseStreamsToSend,
+                  /*cork_metadata=*/true, /*first_write_async=*/false);
+  test.Await();
+  // Make sure that the server interceptors were not notified of a cancel
+  if (GetParam().use_interceptors) {
+    EXPECT_EQ(0, DummyInterceptor::GetNumTimesCancel());
+  }
+}
+
+TEST_P(ClientCallbackEnd2endTest, BidiStreamCorkedFirstWriteAsync) {
+  MAYBE_SKIP_TEST;
+  ResetStub();
+  BidiClient test(stub_.get(), DO_NOT_CANCEL,
+                  kServerDefaultResponseStreamsToSend,
+                  /*cork_metadata=*/true, /*first_write_async=*/true);
   test.Await();
   // Make sure that the server interceptors were not notified of a cancel
   if (GetParam().use_interceptors) {
@@ -1248,8 +1318,10 @@ TEST_P(ClientCallbackEnd2endTest, BidiStream) {
 TEST_P(ClientCallbackEnd2endTest, ClientCancelsBidiStream) {
   MAYBE_SKIP_TEST;
   ResetStub();
-  BidiClient test{stub_.get(), DO_NOT_CANCEL,
-                  kServerDefaultResponseStreamsToSend, ClientCancelInfo{2}};
+  BidiClient test(stub_.get(), DO_NOT_CANCEL,
+                  kServerDefaultResponseStreamsToSend,
+                  /*cork_metadata=*/false, /*first_write_async=*/false,
+                  ClientCancelInfo(2));
   test.Await();
   // Make sure that the server interceptors were notified of a cancel
   if (GetParam().use_interceptors) {
@@ -1261,7 +1333,8 @@ TEST_P(ClientCallbackEnd2endTest, ClientCancelsBidiStream) {
 TEST_P(ClientCallbackEnd2endTest, BidiStreamServerCancelBefore) {
   MAYBE_SKIP_TEST;
   ResetStub();
-  BidiClient test{stub_.get(), CANCEL_BEFORE_PROCESSING, 2};
+  BidiClient test(stub_.get(), CANCEL_BEFORE_PROCESSING, /*num_msgs_to_send=*/2,
+                  /*cork_metadata=*/false, /*first_write_async=*/false);
   test.Await();
   // Make sure that the server interceptors were notified
   if (GetParam().use_interceptors) {
@@ -1274,7 +1347,9 @@ TEST_P(ClientCallbackEnd2endTest, BidiStreamServerCancelBefore) {
 TEST_P(ClientCallbackEnd2endTest, BidiStreamServerCancelDuring) {
   MAYBE_SKIP_TEST;
   ResetStub();
-  BidiClient test{stub_.get(), CANCEL_DURING_PROCESSING, 10};
+  BidiClient test(stub_.get(), CANCEL_DURING_PROCESSING,
+                  /*num_msgs_to_send=*/10, /*cork_metadata=*/false,
+                  /*first_write_async=*/false);
   test.Await();
   // Make sure that the server interceptors were notified
   if (GetParam().use_interceptors) {
@@ -1287,7 +1362,8 @@ TEST_P(ClientCallbackEnd2endTest, BidiStreamServerCancelDuring) {
 TEST_P(ClientCallbackEnd2endTest, BidiStreamServerCancelAfter) {
   MAYBE_SKIP_TEST;
   ResetStub();
-  BidiClient test{stub_.get(), CANCEL_AFTER_PROCESSING, 5};
+  BidiClient test(stub_.get(), CANCEL_AFTER_PROCESSING, /*num_msgs_to_send=*/5,
+                  /*cork_metadata=*/false, /*first_write_async=*/false);
   test.Await();
   // Make sure that the server interceptors were notified
   if (GetParam().use_interceptors) {