Memory leak of TString.tensorOf(Shape shape, DataBuffer<String> data)

Please make sure that this is a bug. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:bug_template

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04 x86_64): macOS 11.5.2
• TensorFlow installed from (source or binary): Binary
• TensorFlow version (use command below): 2.5
• Java version (i.e., the output of java -version): Java HotSpot(TM) 64-Bit Server VM (build 25.261-b12, mixed mode)
• Java command line flags (e.g., GC parameters): -ea
• Python version (if transferring a model trained in Python): 3.7
• Bazel version (if compiling from source): N/A
• GCC/Compiler version (if compiling from source): N/A
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A
• Tensorflow Java version: 0.3.2

Describe the current behavior
When I create and destroy TString instances repeatedly in a for loop, memory usage grows. In my production environment, where a TF model is served by a java service, it would cause memory usage alert after processing certain amount of requests.

Describe the expected behavior
Memory usage should not grow since created TString instances are closed in each iteration.

Code to reproduce the issue

    @Test
    void testPerformance() throws Exception {
        // Some code to load a model but never used it in this test case. Otherwise, the following code would exit randomly. See https://github.com/tensorflow/java/issues/370 .
        final String[] result = new String[] {"a", "b", "c", "d", "e", "f", "g", "h", "i", "c", "c", "c"};
        for (int i = 0; i < 65536; ++i) {
            try (final TString tstring = TString.tensorOf(Shape.of(1, result.length), DataBuffers.ofObjects(result))) {
               // this block is empty.
            }
        }
  }

Other info / logs
4 iterations

8192 iterations

65536 iterations

The above are IntelliJ IDEA's memory profiling screenshots. The differences are as following:

• The first was recorded when the for loop runs 4 iterations (; i < 4; ...). The TString.tensorOf calls only use 2% memory of its parent call.
• The second wass recorded when the for loop runs 8192 iterations (; i < 8192; ...). The TString.tensorOf calls use 49.05% memory of its parent call.
• The third was recorded when the for loop runs 65536 iterations (; i < 65536; ...). The TString.tensorOf calls use up to 83.56% memory of its parent call.

Start by running the provided Java reproduction around TString.tensorOf(Shape, DataBuffer) and DataBuffers.ofObjects(result), then inspect how TString.close() releases resources. Compare memory behavior across the 4, 8192, and 65536 iteration cases. Done means repeatedly created and closed TString instances no longer cause memory usage to grow unexpectedly.