How does it compare with Kotlin for backend development, which inherits all of the JVM ecosystem?
It's the same with F#. F# used to be ahead of C#, but now every time there's a new C# language feature it's F# that has to adjust or deal with friction in some other way. Discriminated unions are the next big thing that will require F# to put up with C#.
As a big C# proponent, my gut tells me the JVM probably has the edge here, but it's not a clean sweep. The CLR has areas where it excels. I would also freely admit that the Java ecosystem is far larger and more mature. The NuGet ecosystem is no slouch, however.
But if I were a gambling man, I would bet that AOT compilation will become much more widespread in the .NET ecosystem while Oracle will ensure that you have to pay the piper for GraalVM to do anything interesting. Call it a hunch.
Where .NET shines is by providing a performance ceiling completely unmatched by neither Java nor Go nor any other GC-based language. Performance ceiling in .NET sits every so slightly below Rust and C, in often an indistinguishable way. It is a very nice experience to write systems code in. It won't give you the same safety for writing concurrent code Rust does, but it is so much more productive than dealing with C or C++, especially when it comes to portable builds and tooling. Of course, it supports only a tiny fraction of platforms compared to C. But your target is more likely to be a server or a consumer device or a comparatively beefy Raspberry PI, and .NET works really well on all of those.
How does it do this, exactly? What does .NET have that is better than Java for performant code?
CIL bytecode, besides what JVM bytecode exposes, provides much lower level access. As a result, C# and Java are languages of different categories and weight classes completely as of 2025. C# is a proper systems programming language in all the ways Java is not. JVM bytecode is also a more difficult to optimize target because all calls are virtual by default and type information is erased. OpenJDK HotSpot has to perform a lot of work just to get to the baseline where CIL starts at, where you have to explicitly make a call virtual and where type information propagates through generic type arguments. OpenJDK used to have a much more powerful compiler but .NET has closed the gap in almost every area, and in other it has surpassed it, or will surpass in the upcoming release(s). It also has a world of optimizations for structs and struct generics which are monomorphized like in Rust, something OpenJDK might only get to in the future. As I said in my previous comment, performance ceiling of .NET is at approximately the level of C/C++/Rust/Zig. Somewhat lower due to compiler limitations but the gap is small enough for it to be easily competitive.
https://benchmarksgame-team.pages.debian.net/benchmarksgame/... is a good demonstration of performance difference in optimized code in these two languages (note that on <1s execution time benchmarks the startup impact also works against both, so you could look at the comparison between C# AOT and Go to get another picture)
That would be 5 measurements out of 100?
With the tiny tiny benchmarks game programs startup and warmup and shutdown costs must effect measurements longer than 1s -- but less so.
https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
Here reverse-complement and spectral-norm have >25% difference. When you look at the main comparison table for either, it might seem that C# lags behind natively compiled languages, but if you look at the NAOT numbers - it is right there next to them.
So I'm putting a disclaimer because of it.
https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
Not really. You don't have the same level of control over memory in managed languages
- structs, with auto, sequential and explicit layout, with complex SROA and promotion handling
- stack-allocated buffers, unsafe fixed and inline arrays, even managed objects can have their layout configured
- monomorphized struct generics compiled the same way they do in Rust
- raw and byref pointers with pointer arithmetic
- portable and platform-specific SIMD, platform-specific intrinsics
- zero-cost interop which enables calls into malloc/free at the cost of C/C++ (you do not actually need this - there is a managed reimplementation of mimalloc which is fully competitive with the original)
- static linking with native dependencies when using nativeaot
Here's a library that does this in C in case you are wondering: https://github.com/rxi/microui.
// Please, just allocate a normal array instead, there is no benefit to this if it lives throughout the whole application lifetime
var bytes = (stackalloc byte[512 * 1024]);
// or, but, please, at least use a span instead of C nonsense of tracking offsets by hand
// it's not 1980 anymore
var ptr = stackalloc byte[512 * 1024]; // byte*
// or
[InlineArray(256 * 1024)]
struct CommandList { byte _; }
var commands = new CommandList();
// or
unsafe struct State
{
public fixed byte CommandList[256 * 1024];
}
To avoid running into stack space limitations, you can place a fixed buffer or any of these large arrays into a static field. It gets a fixed static offset in the memory.
C# goes much further and also provides zero-cost abstractions via struct generics with interface constraints. Or without - you can also apply pattern matching against a generic T struct and it will be zero-cost and evaluated at compilation (be it with RyuJIT or ILC (IL AOT Compiler)).
For example, in Go, the development productivity is great, but I'm not so sure about feature development velocity. There are a ton of HTTP libraries, but it's a barren wasteland when it comes to Auth solutions and you have to rely on a separate service which unnecessarily complicates the infrastructure. Need to quickly put together an application that supports enterprise OAuth? tough luck
I have to admit that I haven't done much with the interweb side. Truth be told, the few times I've used Asp.Net, I used KeyCloak as the authentication provider, and that's Java-based, lol.
That being said (and damn you for making me link to ms docs yet another time today), there is a built in provider, but my recollection is it did not support OAuth2 or JWT: https://learn.microsoft.com/en-us/aspnet/core/security/?view...
It may not suit your needs, but maybe that high-level doc can help you drill down quickly and not waste too much time on it.
I will say that I do always find it amazing how large sites can be sucessfully operated without too much fuss on the platform. Stack Overflow was notorious for being a very high traffic site that ran on a small cluster of machines and was all done in Asp.Net (might have been the old Windows framework, however).
I would like to add that I've been working with Entity Framework (EFCore) for the past year, and I've found it to provide velocity on that front. Never was a big ORM fan, but I can definitely see the use cases now.
????
That's super easy. I added support for Clerk's OAuth to our app in literally 10 minutes. It was as easy as:
> token, err := jwt.ParseString(sessionToken, jwt.WithKeySet(r.keySet), jwt.WithAudience(r.audience))
JWT is from `github.com/lestrrat-go/jwx`
I'm not sure about SAML, but at this point it's probably best to not even touch it.
What else do you need?
That being said, Java has recently released Project Loom, but it's poorly integrated and often causes subtle issues. It'll require a couple more years to mature.
C# doesn't have anything like this for now.
https://hez2010.github.io/async-runtimes-benchmarks-2024/tak...
This overhead is indeed real, but it is basically immaterial. If you're using that many goroutines, you are going to be CPU-bound unless it's a very special type of workload like a WebSocket dispatcher.
In return, you get _normal_ stack traces, and a normal debugging experience. Not a callback hell of async/await. Also no "colored functions" nonsense.
Coroutines make sense only for something small and trivial, like the classic tree iterator. And Go now has that: https://go.dev/blog/range-functions
Whatever language you have in mind, it isn't C#. Because all those work perfectly fine there. Please also do a cursory read of what coroutines (stackful vs stackless) in the context of concurrency are. And in Go, all functions are colored, often in ambiguous way due to poor culture of not passing the context where it matters, and goroutines panicking in dependencies that cause uncatchable application crashes. Never change, Go community.
Stackful coroutines have their pros but they are not a better choice. Rust did not and will not adopt them nor any other serious systems programming language will.
You don’t have to follow noisy style you often see - tasks compose nicely and express deferred operations incredibly well in a way that will not blow up on you in a surprising way.
There is an ongoing Async2 project which will massively reduce task state machine overhead further by only ever paying for it in the code that actually suspends rather than the code that just forwards the calls or doesn’t. It will likely land as preview in .NET 10 and as full release in .NET 11.
You seem knowledgeable on the matter, care to share some resources that might help me grok the differences?
Hard disagree. Coroutines are absolutely useless for anything non-trivial, as debugging them becomes a total hell. They are still a callback hell, just with a lot of sugary goo slathered on it.
Coroutines also result in the "colored function" problem, that is fundamental for them.
Meanwhile, Go lightweight threads just work. Debugging is simple, and you don't have to think about spooky actions at a distance from event loops that dispatch coroutines.