Java Native Interface

When you say you're using in-memory data stores, are any of those using ...

A lot of data stores use off-cheap memory, right? So the memory is held, it's malloced by some JNI C++ code, right? There's just a malloc, 10 gigs, and then the JVM manually puts things into this native buffer off to the side.

Is that what you're doing, or is that actual Java objects on the JVM

in the deep? Well, first of all, we have a whole bunch of different data stores specialized for specific use cases.

And in C, there's a popular strlen function that's going to give you the length of a string, but it returns an unsigned type.

And if you look in JNI, then the array length is actually given to you in a thing that they call a Jsize, which is an unsigned integer.

Mm-hmm.

And so clearly one set of people, when Java was being created, were thinking, well, it doesn't really make sense to have negatively sized arrays.

Having lightweight threading, as in virtual threads in Loom, was another.

Having access to native libraries without going through JNI was another thing.

One by one, each of these things that I presented, and I think 2013 or so, has come true.

It's come to the JVM.

So what everybody has been working towards over the years is a way to integrate these managed runtimes with their own heaps and their own garbage collectors with native libraries that have basically their own memory model, their own way of doing method calls.

And so we've had things like FFA on the JVM.

We've had JNA for a long time, Java Native Access, which basically just has a little stub of JNI code to do the calls for you to the C logic.

We have our own ecosystem in JRuby called JNR, Java Native Runtime.

I built a zero-copy Fog AI node without Python, and it hurt.

A deep dive into building Fog AI, ditching Python for JNI and C++ to run Qwen3-VLON orange Pi4 with 60x lower latency for industrial edge AI.

Ninety-nine.

Seven mistakes to avoid when hiring an app development company.

That's not an issue.

There's a runtime pretty much up there, like RT and XGTOS, where you can use it, right? But when we talk about the maybe JNI models, which is present right now, people don't know how to make it very smaller, which is called in technical terms quantization, right? Or when we talk about pruning techniques or those kinds of techniques, right? So a lot of folks didn't know that.

And a lot of folks don't know what is their use case and why they are using that particular model for that, right? So sometimes people use general model to implement some specific use cases.

If they can fine tune that particular model, and obviously if you're fine tuning a model, you are making the model more smaller also, apart from the quantization work and all.