// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromScalar(
    TF_DataType data_type, void* data, size_t len, TF_Status* status);

// Specify the server_def that enables collective ops.

   * been shutdown).
   *
   * <p>Although all tasks are immediately executed in the thread that submitted the task, this
   * {@code ExecutorService} imposes a small locking overhead on each task submission in order to
   * implement shutdown and termination behavior.
   *
   * <p>The implementation deviates from the {@code ExecutorService} specification with regards to

But if you are certain that the content that you are returning is **serializable with JSON**, you can pass it directly to the response class and avoid the extra overhead that FastAPI would have by passing your return content through the `jsonable_encoder` before passing it to the response class.

```Python hl_lines="2  7"
{!../../docs_src/custom_response/tutorial001b.py!}
```

/// info

                                                    TF_Status* status);
// Get an attribute type given an op name; a fusion of TFE_NewOp and
// TFE_OpGetAttrType for use from Python without the overhead of the individual
// calls and memory management of TFE_Op.
TF_CAPI_EXPORT extern TF_AttrType TFE_OpNameGetAttrType(
    TFE_Context* ctx, const char* op_or_function_name, const char* attr_name,

The *Secure Channel* splits the object content into chunks of a fixed size of `65536` bytes. The last chunk may be smaller to avoid adding additional overhead and is treated specially to prevent truncation attacks. The nonce value is 96 bits long and generated randomly per object / multi-part part. The *Secure Channel* supports plaintexts up to `65536 * 2^32 = 256 TiB`.

#### Randomness

 * form}), the memory requirement is 8*N bytes for the copy of the dataset plus an overhead which is
 * independent of N (but depends on the quantiles being computed). When calling {@link
 * ScaleAndIndex#computeInPlace computeInPlace} (in {@linkplain ScaleAndIndexes#computeInPlace
 * either form}), only the overhead is required. The number of object allocations is independent of
 * N in both cases.
 *

 *       allocating garbage).
 * </ul>
 *
 * <p>Disadvantages compared to {@code long[]}:
 *
 * <ul>
 *   <li>Memory footprint has a fixed overhead (about 24 bytes per instance).
 *   <li><i>Some</i> construction use cases force the data to be copied (though several construction
 *       APIs are offered that don't).

 *       allocating garbage).
 * </ul>
 *
 * <p>Disadvantages compared to {@code int[]}:
 *
 * <ul>
 *   <li>Memory footprint has a fixed overhead (about 24 bytes per instance).
 *   <li><i>Some</i> construction use cases force the data to be copied (though several construction
 *       APIs are offered that don't).

   * non-static, and is initialized/removed frequently - this causes churn in the Thread specific
   * hashmaps. Using a static ThreadLocal to avoid that overhead would mean that different
   * ExecutionSequencer objects interfere with each other, which would be undesirable, in addition
   * to increasing the memory footprint of every thread that interacted with it. In order to release

### Multi destination replication

The replication design for multiple sites works in a similar manner as described above for two site scenario. However there are some
important exceptions.