Like the article says, fan-out is a key design characteristic. There are "redis streams" now but they didn't exist back then. The durability story and cluster stories aren't as good either, I believe, so they can probably take you so far but won't be as generally suitable depending on where your system goes in the future. There are also things like RedPanda that speak Kafka w/o the Java.
However, if you CAN run on a single node w/o worrying about partitioning, you should do that as long as you can get away with it. Once you add multiple partitions ordering becomes hard to reason about and while there are things like message keys to address that, they have limitations and can lead to hotspotting and scaling bottlenecks.
But the push/pop based systems also aren't going to give you at-least-once guarantees (looks like Redis at least has a "pop+push" thing to move to a DIFFERENT list that a single consumer would manage but that seems like it gets hairy for scaling out even a little bit...).
I'm curious, what exactly feels bloated about Java? I don't feel like the Java language or runtime are particularly bloated, so I'm guessing you're referring to some practices/principles that you often see around Java software?
Yes yes I’m sure there are exceptions somewhere but I’ve been reading Java fans using benchmarks to try to convince me that I can’t tell which programs on my computer are Java just by looking for the weirdly slow ones, when I in fact very much could, for 25ish years.
Java programs have a feel and it’s “stuttery resource hog”. Whatever may be possible with the platform, that’s the real-world experience.
I haven't worked with too much Java, but I suspect that the distaste many have for it is due to its wide adoption by large organizations and the obfuscating "dressed up" tendency of the coding idioms used in large organizations.
The runtime isn't inherently slow, but maybe it's easier to write slow programs in Java.
jokes aside, we got a shift in the industry where many java programs were replaced by electron-like programs which now take 20x memory
While for typical backend situations, reference counting has a crazy high throughput overhead, doing atomic inc/decs left and right, that instantly trashes any kind of cache, and does it in the mutator thread that would do the actual work, for the negligible benefit of using less memory. Meanwhile a tracing GC can do (almost) all its work in another thread, not slowing down the actually important business task, and with generational GCs cleaning up is basically a no-op (of just saying that this region can now be reused).
It's a tradeoff as everything in IT.
Also, iPhone CPUs are always a generation ahead, than any android CPU, if not more. So it's not really Apples to oranges.
* languages like c++ and Rust simply don’t allocate as much as Java, instead using value types. Even C# is better here with value types being better integrated.
* languages like c++ and Rust do not force atomic reference counting. Rust even offers non atomic ref counting in the standard library. You also only need to atomic increment / decrement when ownership is being transferred to a thread - that isn’t quite as common depending on the structure of your code. Even swift doesn’t do too badly here because of the combination of compiler being able to prove the permission of eliding the need for reference counting altogether and offering escape hatches of data types that don’t need it.
* c++, Rust, and Swift can access lower level capabilities (eg SIMD and atomics) that let them get significantly higher throughput.
* Java’s memory model implies and requires the JVM to insert atomic accesses all over the place you wouldn’t expect (eg reading an integer field of a class is an atomic read and writing it is an atomic write). This is going to absolutely swamp any advantage of the GC. Additionally, a lot of Java code declares methods synchronized which requires taking a “global” lock on the object which is expensive and pessimistic for performance as compared with the fine-grained access other languages offer.
* there’s lots of research into ways of offering atomic reference counts more cheaply (called biased RC) which can safely avoid needing to do an atomic operation in places completely transparently and safely provided the conditions are met .
I’ve yet to see a Java program that actually gets higher throughput than Rust so the theoretical performance advantage you claim doesn’t appear to manifest in practice.
AFAIK that doesn’t really happen. They won’t insert atomic accesses anywhere on real hardware because the cpu is capable of doing that atomically anyway.
> Additionally, a lot of Java code declares methods synchronized which requires taking a “global” lock on the object which is expensive and pessimistic for performance as compared with the fine-grained access other languages offer.
What does this have to do with anything? Concurrency requires locks. Arc<T> is a global lock on references. “A lot” of Java objects don’t use synchronized. I’d even bet that 95-99% of them don’t.
Concurrency does not require locks. There’s entire classes of lock free and wait free algorithms. Arc<T> is also not a lock - it uses atomics to manage the reference counts and no operation on an Arc needs to wait on a lock (it is a lock-free container).
> “A lot” of Java objects don’t use synchronized. I’d even bet that 95-99% of them don’t.
Almost all objects that are used in a concurrent context will likely feature synchronized, at least historically. That’s why Hashtable was split into HashMap (unsynchronized) and ConcurrentHashMap (no longer using synchronized). Thats why you have StringBuffer which was redone into StringBuilder.
> and to implement a ConcurrentHashMap in Rust you would still need to lock
There’s many ways to implement concurrency safe hashmaps (if you explicitly needs such a data structure as the synchronization mechanism) without locks. Notably RCU is such a mechanism (really neat mechanism developed for the kernel although not super user friendly yet or common in userspace) and there are also generational garbage techniques available (kind of similar to tracing GC conceptually but implemented just for a single data structure). A common and popular crate in Rust for this is DashMap which doesn’t use locks and is a concurrency safe hashmap.
Still not in the standard library. The only way in Rust is to use a global lock around map. Seems to be worse than the situation in Java. You could implement the same thing and use a third party library in Java too. So your original point of "everything uses a global lock" is "overly strong"
The strengths and weaknesses of the standard library aren’t relevant. But if we’re going there, the reason they’re not in the Rust standard library is likely in practice concurrent data structures are an anti pattern - putting a lock around a data structure doesn’t suddenly solve higher order race conditions which is something a lot of Java programmers seem to believe because the standard library encourages this kind of thinking.
As for “my comment” about “global lock” (your words not mine), it’s that the implicit lock that’s available on every object is a bad idea for highly concurrent code (not to mention the implicit overhead that implies for every part of the object graph regardless of it being needed anywhere). Don’t get me wrong - Java took a valiant effort to define a solid memory model for concurrency when the field was still young. Many of the ideas didn’t pan out and are antipatterns these days for high performing code. Of course none of that pertains to the original point of the conversation - tracing GCs have significantly more overhead in practice because they’re very difficult to be opt in, carry quite a penalty if not, Rc/Arc is much better as it’s possible to do opt-in when you need shared ownership (which isn’t always), and in practice loops don’t come up often enough to matter and when they do there’s still solutions. In other words tracing GCs drop huge amounts of performance on the floor and you can read all the comments to see how the claims are “it’s more efficient than Rc”, or “performance is free” or even “it doesn’t matter because the programmer is more efficient”. I’d buy the efficiency argument when the only alternative was C/C++ and came with serious memory safety baggage, but not any of the others and memory safety without sacrificing performance of C++ in my view is a solved problem with Rust.
Of course with manual memory management you may be able to write more efficient programs, though it is not a given, and comes at the price of a more complicated and less flexible programming model. At least with Rust, it is actually memory safe, unlike c++.
- ref counting still has worse throughout than a tracing GC, even if it is single-threaded, and doesn't have to use atomic instructions. This may or may not matter, I'm not claiming it's worse, especially when used very rarely as is the case with typical c++/rust programs.
> You also only need to atomic increment / decrement when ownership is being transferred to a thread
Java can also do on-stack replacement.. sometimes.
- regarding lower level capabilities, java does have an experimental Vector API for simd. Atomics are readily available in the language.
- Java's memory model only requires 32-bit writes to be "atomic" (though in actuality the only requirement is to not tear - there is no happens before relation in the general case, and that's what is expensive), though in practice 64-bit is also atomic, both of which are free on modern hardware. Field acces is not different from what rust or c++ does, AFAIK in the general case. And `synchronized` is only used when needed - it's just syntactic convenience. This depends on the algorithm at hand, there is no difference between the same algorithm written in rust/c++ vs java from this perspective. If it's lockless, it will be lockless in Java as well. If it's not, than all of them will have to add a lock.
The point is not that manual memory can't be faster/more efficient. It's that it is not free, and comes at a non-trivial extra effort on developers side, which is not even a one-time thing, but applies for the lifetime of the program.
That’s a bold claim to make that doesn’t seem to actually be true from my experience. Your 5ghz CPU can probably do ~20 billion non atomic reference adjustments whereas your GC system has to have atomics all over the place or it won’t work and atomics have parasitic performance on unrelated code due to bus locks and whatnot.
> Java can also do on-stack replacement.. sometimes
That’s not what this is. It’s called hybrid RC and it applies always provided you follow the rules.
> The point is not that manual memory can't be faster/more efficient. It's that it is not free, and comes at a non-trivial extra effort on developers side, which is not even a one-time thing, but applies for the lifetime of the program.
The argument here is not about developer productivity - the specific claim is that the Java GC lets you write higher throughput code than you would get with Rust or C++. That just isn’t true so you end up sacrificing throughput AND latency AND peak memory usage. You may not care and are fine with that tradeoff, but claiming you’re not making that tradeoff is not based on the facts.
No, that has never been the specific claim - you can always write more efficient code with manual memory management, given enough time, effort and skill. I wasn't even the one who brought up c++ and rust. Like literally I write this twice in my comment.
What I'm talking about is reference counting as a GC technique vs tracing as a GC technique, all else being equal - it would be idiotic to compare these two if no other "variable" is fixed. (Oh and I didn't even mention the circular references problem, which means you basically have to add a tracing step either-way unless you restrict your language so that it can't express circular stuff).
As for the atomic part, sure, if all it would do is non-atomic increments then CPUs would be plenty happy. And you are right that depending on how the tracing GC is implemented, it will have a few atomic instructions. What you may miss is how often each run. On almost every access, vs every once in a while on a human timescale. Your OS scheduler will also occasionally trash the performance of your thread. But this is the actually apples to oranges comparison, and both techniques can do plenty of tweaks to hide certain tradeoffs, at the price of something else.
And I also mention that the above triad of time, skill and effort is not a given and is definitely not free.
As for “skill” this is one thing that’s super hard to optimize for. All I can do is point to existence proofs that there’s no mainstream operating system, browser or other piece of high performance code written in Java and it’s all primarily C/C++ with some assembly with Rust starting to take over the C/C++ bits. And at the point where you’re relegating Java to being “business” logic, there’s plenty of languages that are better suited for that in terms of ergonomics.
Java is used in HFT (well, there are two types of "high frequency", one where general purpose CPUs are already too slow, where it obviously doesn't apply (neither do rust or c++)) - but sure, I wouldn't write a runtime or other piece of code in Java where absolute control over the hardware is required. But that's a small niche only. What about large distributed systems/algorithms? Why is Java over-represented in this niche (e.g. Kafka, Elasticsearch, etc)?
> And at the point where you’re relegating Java to being “business” logic, there’s plenty of languages that are better suited for that in terms of ergonomics.
That's subjective.
The JVM tends to hold onto memory in order to make things faster when it does wind up needing that memory for actual stuff. However, how much it holds on to, how the GC is setup, etc are all tunable parameters. Further, if it's holding onto memory that's not being used, these are prime candidates to be stored in virtual memory which is effectively free.
Is C++ bloated because of the memory Chrome uses?
I’ve never seen another basic tech used to develop other programs that’s so consistently obvious from its high resource use and slowness, aside from the modern web platform (Chrome, as you put it). It was even more obvious back when we had slower machines, of course, but Java still stands out. It may be able to calculate digits of Pi in a tight loop about as fast as C, but real programs are bloated and slow.
Especially that like half of the web runs on Java, you just have absolutely no idea when it silently does its job perfectly.
Source?
W3 seems to think its more like ~5%
From the website.
And 76% of these websites is PHP, which seems to mean.. they can determine PHP more easily for a website (nonetheless, there are indeed a lot of WordPress sites, but not this amount).
With over 15 years of professional experience since then, my perspective has shifted: Java demonstrates its strength when stability, performance, and scalability are required (e.g. bloody enterprise)
A common misconception comes from superficial benchmarking. Many focus solely on memory consumption, which often provides a distorted picture of actual system efficiency.
I can point to EU-scale platforms that have reliably served over 100 million users for more than a decade without significant issues. The bottleneck is rarely the language itself, it is the depth of the team’s experience.
When other languages can do the same thing with an order of magnitude less RAM, any other efficencies in the system tend to be overshadowed by that and be the sticking point in peoples memories.
You may argue that holding on to this extra memory makes subsequent calls and reads quicker etc, but in my experience generally people are willing to sacrifice milliseconds to gain gigabytes of memory.
Nonetheless, tracing GCs do have some memory overhead in exchange for better throughput. This is basically the same concept as using a buffer.
-----
And can you tell which of these websites use Java from "the feel"? AWS cloud infra, a significant chunk of Google, Apple's backends, Alibaba, Netflix?
The way most Java code is written is terrible Enterprise factory factory factory.
Worse latency every ten minutes tends to be fine.
Go's GC is tuned more for latency at the expense of throughput (not sure if it still applies, but Go was quite literally stopping the "business" mutator threads when utilisation got higher to be able to keep up with the load - Java's default GC is tuned for a more balanced approach, but it can deliver it at very high congestion rates as well. Plus it has a low-latency focused GC which has much better latency guarantees, and it trades off some throughput in a consistent manner, so you can choose what fits best). The reason it might sometimes be more efficient than Java is simply value types - it doesn't create as much garbage, so doesn't need as good a GC in certain settings.
Rust code can indeed be better at both metrics for a particular application, but it is not automatically true, e.g. if the requirements have funny lifetimes and you put a bunch of ARC's, then you might actually end up worse than a modern tracing GC could do. Also, future changes to the lifetimes may be more expensive (even though the compiler will guide you, you still have to make a lot of recursive changes all across the codebase, even if it might be a local change only in, say, Java), so for often changing requirements like most business software, it may not be the best choice (even though I absolutely love Rust).
Such as? The only area where you have to "drop" features is high-frequency trading, where they often want to reach a steady-state for the trading interval with absolutely no allocations. But for HFT you would have to do serious tradeoffs for every language.
In my experience, vanilla java is more than fine for almost every application - you might just benchmark your code and maybe add some int arrays over an Integer list, but Java's GC is an absolute beast, you don't have to baby it at all.
The reason is quite well known. Supporting multiple languages is a cost. If you only have to support one language, everything is simpler and cheaper.
With Java, you can write elegant code these days, rely on ZGC, not really worry too much about GC and get excellent performance with quick development cycles for most of your use cases. Then with the same language and often in the same repo (monorepo is great) you can write smarter code for your hot path in a GC free manner and get phenomenal performance.
And you get that with only having one build system, one CI pipeline, one deployment system, some amazing profiling and monitoring tooling, a bunch of shared utility code that you don't have to duplicate, and a lot more benefits.
That's the reason to choose Java.
Of course, if you're truly into HFT space, then they'll be writing in C, C++ or on FPGAs.
Maybe I have been lucky, or that the practice is more common in certain countries or eco systems? Java has been a very productive language for me, and the code has been far from the forced pattern usage that I have read horror stories about.
Depending on how much memory is used by the Thread stack (presumably 1M-512K by default, allegedly 128K with Alpine base images) that's your 1G-500M heap space usage improvement right off the bat.
The migration from JDK17 to JDK21 was uneventful in production. The only issue is limited monitoring as a thread dump will not show most virtual threads and the micrometer metrics will not even collect the total number of active virtual threads. It's supposed to work better in JDK24.
The Spring Framework directly supports virtual threads with "spring.threads.virtual.enabled=true" but I haven't tried it to comment.
Everything.
Why do you think Kubernetes is NOT written in Java?
Golang has little to distinguish itself technically. It has a more modern std lib (for now) and isn't Oracle.
Which aren't trivial, but they aren't Trump cards.
Nope.
None of what you said are any of the reasons given that it WAS written in Java already [0] but rewrote it all in Go explicitly because of its performance, concurrency and single binary distribution characteristics.
Those were enough technical advantages to abandon any thought of a production-grade version of k8s in Java.
[0] https://archive.fosdem.org/2019/schedule/event/kubernetesclu...
It seems to me that perhaps it wasn’t the languages fault but the authors.
Because someone wanted a new, shiny toy.
https://docs.spring.io/spring-framework/docs/2.5.x/javadoc-a...
They hog RAM, are slow, and are a bitch to configure.
Kafka is used because the Java folks don't want to learn something new due to job security, even though there are faster and compatible alternatives that exist today.
Rather use Redpanda, than continue to use Kafka and then complain about how resource intensive it is alongside zookeeper and all the circus the comes with it and make AWS smile as you're losing hundreds of thousands a month.
1: https://cwiki.apache.org/confluence/display/kafka/kip-500:+r...
https://www.confluent.io/blog/zookeeper-to-kraft-with-conflu...