[BUG] [Java] Memory leak due to table writers never being destroyed after done writing

[ttnghia]

For all kinds of table writers (Java code: Table.java, JNI code: TableJni.cpp), the native C++ instances are never destroyed after calling close() in the corresponding Java classes. For example:

@Override
    public void close() throws CudfException {
      if (writerHandle != 0) {
        writeParquetEnd(writerHandle);
      }
      writerHandle = 0;

And here is the C++ function writeParquetEnd:

  try {
    cudf::jni::auto_set_device(env);
    state->writer->close();
  }

Note that state is an unmanaged pointer that corresponds to writerHandle in Java and stores writer as a unique_ptr. So the writer instance is destroyed automatically when state is destroyed, however, state is never destroyed.

When writeParquetEnd is called, the C++ code only calls the member close() function of the writer, never calls delete on the state instance. After this, the Java code simply leaves that native instance dangled.