// scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph

  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution
  // fails on any device in order to cancel the others.
  void StartExecute(TFE_Context* context, const char* operation_name,

// Deletes the GCS directory marker if it exists.
void DeleteDir(const TF_Filesystem* filesystem, const char* path,
               TF_Status* status) {
  // A directory is considered empty either if there are no matching objects
  // with the corresponding name prefix or if there is exactly one matching
  // object and it is the directory marker. Therefore we need to retrieve

//     but plugins didn't provide will be set to `nullptr` values and core
//     TensorFlow will know to not call these on behalf of users;
//   * increased security as plugins will not be able to alter function table
//     after loading up. Thus, malicious plugins can't alter functionality to
//     probe for gadgets inside core TensorFlow. We can even protect the area
//     of memory where the copies reside to not allow any more writes to it

  std::vector<std::unique_ptr<Thread>> threads;
  threads.reserve(callers);
  for (int i = 0; i < callers; i++) {
    threads.emplace_back(
        Env::Default()->StartThread({}, "caller", [block_size, &cache, i]() {
          std::vector<char> out;
          TF_EXPECT_OK(
              ReadCache(&cache, "a", i * block_size, block_size, &out));
          std::vector<char> x(block_size, 'x');