Dimitar Makariev’s Post

You’re saying completely dominating but the article names Go as the winner here because Java used 12x the memory and the speed is virtually identical. So, a bit misleading?

Now lets compare numbers of AI libraries, frameworks and connectors adapted for Python and TS vs Java and Go. Whats the point of milliseconds advantage in runtime if your competitors would get all clients, while you would run behind in features, that are already tackled in most opensource libraries / SaaS solutions?

I'm not sure if now is a good time for Java developers. According to the latencies, it seems that Python developers have more time for a good coffee!

There's one footnote in the article that interests me: "It is worth noting that purely minimalist implementations, aggressive JVM tuning, or GraalVM Native Image compilation could yield significantly different resource efficiency results." I have been playing with GraalVM native mode for some time now, and I actually find this statement to be a significant understatement of what could be possible. GraalVM typically offers 100x improvement in startup times and 5x lower memory consumption. So while your charts are simply fantastic, if you wanted a version that blew Go away, GraalVM native would do that for you.

Speaking about environment efficiancy which spend from 500% to 1000% more resources, especially in memory usage, than native binary environments is insanity on galactic level. The biggest infrastructures on the planet are removing and deleting completely Java from usage, while you are embracing it. I feel sad for you. This is profe that today on digital platforms, anyone can write brainless statements and call themselfs “technical” people.

Go is the best in terms of concurrency. It offers short development time, is easy to learn, and uses less memory. Java, on the other hand, has a huge amount of knowledge embedded within the Spring framework. Personally, I prefer to see and understand all the code when I look at a project. With Java, you often need prior experience to start development effectively, and much of the functionality is hidden inside the Spring framework, which can make it feel like there is a lot of “missing” or abstracted code.

I wonder why Rust is excluded? The rmcp crate is great to build MCP servers with, and can't see why it wouldn't beat Java..

Java architect speaking 🙋🏻. Those are good news and we knew Java would catch up performance wise. But main bottleneck and latency usually comes from data APIs (ML models, LLMs etc), so not a big deal. MCP tasks involves heavy LLM usage

Ouh, here we go again. The headlines "Java and Go are 30 times faster" and the key takeaway "Go is ideal for production" are a misrepresentation of context from this article. They are only valid for this specific set of configurations, where Node.js and Python were intentionally placed in a significantly worse position. This is a classic example of "comparing oranges to apples." Instead of tuning each server for maximum performance (production tuning), the developr took the "standard" (Java, Go) and "safe/default" (Node.js, Python) approaches, which led to predictable results. Few examples : - Node.js's "Security-hardened" configuration - this is a massive artificial slowdown. Rebuilding the server for every request isn't a "standard DX" (Developer Experience) feature like for other languages, but an extreme and very rare scenario that devastated performance. - The "Single-worker uvicorn" Python configuration without optimizations. This is a basic, non-production-ready configuration. For Python in production, the standard is a minimum of 4-8 workers (Gunicorn + Uvicorn), which significantly increases throughput. After similar topics everyone can be proud of himself, however we are engineers and we should proud by results.

Those sub-millisecond p99 latencies for Java are definitely impressive for the orchestration layer. However, we need to be realistic: while Java is dominating the MCP server benchmarks, the actual AI inference still happens deep down in C++ and CUDA territories. Reducing the overhead of the "messenger" is great for high-throughput agents, but it doesn't change the fact that Python still holds the keys to the entire research and library ecosystem. Java is winning the "networking" game here, but the "AI" game is much broader than just server latency.