}

  /**
   * Similar to load64, but allows offset + 8 > input.length, padding the result with zeroes. This
   * has to explicitly reverse the order of the bytes as it packs them into the result which makes
   * it slower than the native version.
   *
   * @param input the input bytes
   * @param offset the offset into the array at which to start reading
   * @param length the number of bytes from the input to read

### 3. Failure after conversion
If the conversion is successful, but the generated model is wrong, then state what is wrong:

- Model produces wrong results and/or has lesser accuracy.
- Model produces correct results, but it is slower than expected.

### 4. (optional) RNN conversion support
If converting TF RNN to TFLite fused RNN ops, please prefix [RNN] in the title.

### 5. (optional) Any other info / logs

    return stringWriter.toString();
  }

  /**
   * Returns the stack trace of {@code throwable}, possibly providing slower iteration over the full
   * trace but faster iteration over parts of the trace. Here, "slower" and "faster" are defined in
   * comparison to the normal way to access the stack trace, {@link Throwable#getStackTrace()

    return stringWriter.toString();
  }

  /**
   * Returns the stack trace of {@code throwable}, possibly providing slower iteration over the full
   * trace but faster iteration over parts of the trace. Here, "slower" and "faster" are defined in
   * comparison to the normal way to access the stack trace, {@link Throwable#getStackTrace()

Java serialization is one way to implement this, however, despite its simplicity of implementation, it has several drawbacks:

- **Performance:**
Java's built-in serialization mechanism is often slower compared to other serialization solutions.
This is due to Java's use of reflection and the need to maintain a lot of metadata.

- **Size of Serialized Data:**

* **Starlette**:
    * Will have the next best performance, after Uvicorn. In fact, Starlette uses Uvicorn to run. So, it probably can only get "slower" than Uvicorn by having to execute more code.
    * But it provides you the tools to build simple web applications, with routing based on paths, etc.

				// been seen recently due to hardware issues
				// causes Reads() to go slower than Writes().
				//
				// Send such results anyways as this shall
				// expose a problem underneath.
				if totalPut > throughputHighestPut {
					throughputHighestResults = results
					throughputHighestPut = totalPut
					// let the client see lower value as well
					throughputHighestGet = totalGet
				}
				sendResult()

   */
  static final class SmoothWarmingUp extends SmoothRateLimiter {
    private final long warmupPeriodMicros;
    /**
     * The slope of the line from the stable interval (when permits == 0), to the cold interval
     * (when permits == maxPermits)
     */
    private double slope;

    private double thresholdPermits;
    private double coldFactor;

    SmoothWarmingUp(

         * This avoids out-of-control memory consumption, and it keeps the cache from growing so
         * large that doing the lookup is noticeably slower than redoing the work would be.
         *
         * Ideally we'd have a real eviction policy, but until we see a problem in practice, I hope

* Large payloads.

Only a single connection is ever made between two servers.
Likely this means that this connection will not be able to saturate network bandwidth.
Therefore, using this for large payloads will likely be slower than using a separate connection,
and other connections will be blocked while the large payload is being sent.

## Handlers & Routes

Handlers have a predefined Handler ID.