Tensorflow work_sharder.h and concurrent read and write access

I am adding a custom op in Tensorflow the single-threaded Op is working fine it compiles and runs when imported.

(It also runs using pure-python and multiprocessing library with a shared Array.)

I am defining a shard function:

#include "work_sharder.h"
//Register Op then overwrite of the Compute function then inside
//arg1 and arg2 are const
auto shard = [arg1, arg2,
              &shared_array]
              (int64 start, int64 limit) {
  //Reading (and sometimes passing by value shared_array,   which is an Eigen::Tensor) often outside the start-limit safe zones.


  //Writing in shared_array only between indexes start and limit so no write/write concurrency
  //For instance

   for (int counter=1; counter<hard_limit; counter++){
   if (shared_array(start-counter) == 1){
   shared_array(start)=0;}}

  //Not sure if relevant but I am also using types from std like std::pair.

};
    const DeviceBase::CpuWorkerThreads& worker_threads = *(context->device()->tensorflow_cpu_worker_threads());
    const int64 shard_cost = N;
    Shard(worker_threads.num_threads, worker_threads.workers,
            shared_array.size(), shard_cost, shard);

It compiles perfectly. When running this multi-threaded code I run into 1592 segfaults.

Is it expected behavior because I am at the same time writing and reading in shared_array ? Or should it be handled gracefully by this type of code ?

If this is expected behavior what would be the minimal set of glue code to make it work ?