Merge pull request #284 from karllessard/ndarray-repo

[RFC] New repository for TensorFlow Java NdArray

Author: theadactyl

sigs/jvm/rfcs/20200726-java-ndarray-repository.md

@@ -0,0 +1,44 @@
+# Java NdArray Repository
+| Status        | Proposed       |
+:-------------- |:---------------------------------------------------- |
+| **Author(s)** | Karl Lessard ([email protected]) |
+| **Sponsored** | Joana Carrasqueira ([email protected]), Rajagopal Ananthanarayanan ([email protected]) |
+| **Updated**   | 2020-07-29                                          |
+
+## Objective
+
+Create a new repository under the TensorFlow GitHub organization to host the code of the NdArray Java library used by TensorFlow and 
+maintained by SIG JVM.
+
+## Motivation
+
+SIG JVM has developed a library called [`NdArray`](https://github.com/tensorflow/java/tree/master/ndarray) to improve the 
+support of n-dimensional data structures on the JVM (e.g. tensors in TensorFlow).
+
+While this library is intensively used by the TensorFlow Java bindings, it does not have any dependency to the 
+TensorFlow runtime itself, making it an interesting option for any Java developer interested in fast and efficient I/O accesses 
+to continuous native memory in a n-dimensional space.
+
+To promote the adoption of this library beyond the scope of TensorFlow users, SIG JVM would like to move it into a distinct 
+GitHub repository.
+
+## User Benefit
+
+N-dimensional memory access in Java is poorly supported by the JDK and while there are already a few initiatives involving different 
+parties to standardize such API, it will take time before it gets released to the public. The NdArray Java Library is now a good option for 
+all kind of Java projects and since it does not depend on the TensorFlow runtime, it is very lightweight and portable.
+
+Also, having a distinct repository for the NdArray Java Library development will simplify the task for its collaborators as they will not need
+anymore to checkout and build the whole TensorFlow Java project, a process which tends to be heavy since it also runs a Bazel build 
+of the TensorFlow core sources that can take many hours.
+
+## Design Proposal
+
+The current request is for the creation of the following repository: `/tensorflow/java-ndarray`
+
+This new repository will only host the code of the NdArray Java Library, which will be moved out of `tensorflow/java`. 
+Everything else related to TensorFlow Java depends on the TensorFlow runtime and shall remain under their actual
+repositories.
+
+The NdArray Java Library will now have its own release cycle and group of collaborators, independently from the TensorFlow Java bindings, while
+remaining under the jurisdiction of SIG JVM.