mwArray vs MATLAB Data API - Which API should I use when using MATLAB Compiler SDK?

When developing libraries with MATLAB Compiler SDK, I see most deployment targets generally have two paths for handling data exchange: the legacy mwArray approach and the modern MATLAB Data API (MDA) approach. Which one should I use?

Answer 1

[MathWorks Support Team]

For new development—especially if you want the best performance —we strongly recommend using MDA and particularly its strongly-typed approach.

Why choose MDA over mwArray?

mwArray was built on older technology and requires manual memory management and explicit data casting. In contrast, MDA (introduced for C++ in R2018a and expanded to .NET/Java/Python since) offers several modern advantages:

• Zero/Minimal Data Copying: MDA uses a shared memory model with copy-on-write semantics. This significantly reduces the memory footprint and latency when passing large matrices between your application and the MATLAB Runtime.
• Thread Safety: Unlike mwArray, MDA is designed for multi-threaded environments, allowing you to call compiled functions across different threads without complex locking.
• Native Language Experience: Using the Strongly Typed approach with MDA, you don't need to wrap your data in MATLAB-specific classes like MWNumericArray. You simply pass native C++ vectors, C# arrays, or Java primitives.
• Simplified Lifecycle: Modern APIs utilize smart pointers (C++) and native garbage collection (.NET/Java/Python) to handle the lifecycle of the MATLAB Runtime (and the libraries used with it) automatically.

Target Language	mwArray / MWArray	MDA
C++	Effort: Manual LibInitialize/Terminate and manual mwArray memory management. Differences: requires deep copies for complex data; not thread-safe.	Effort: Automated lifecycle; uses standard C++11+ types via generateCPP. Differences: Zero-copy/Copy-on-Write; thread-safe; full compile-time type checking; runtime instance can run in-process or out of process, applications can run synchronously or asynchronously.
C# / .NET	Effort: Heavy casting to MWArray types; must call .Dispose() on every array to avoid leaks. Differences: marshaling overhead; type errors only caught at runtime.	Effort: Pass native types (double[], string) directly; memory managed by Garbage Collection. Differences: IntelliSense support for method signatures; faster marshaling; asynchronous call support; runtime instance can run in-process or out of process, applications can run synchronously or asynchronously. Support for cross-platform (Windows, macOS, Linux) development and deployment.
Java	Effort: Explicitly wrap data in MWNumericArray etc.; requires manual .dispose() calls. Differences: High object overhead; slower data transfer due to heavy wrapping.	Effort: Uses standard Java objects/primitives; lifecycle handled by JVM/Runtime. Differences: Cleaner API; better memory efficiency; native support for structs and cells.
Python	Has always used MDA with all its associated benefits