One year after switching from Java to Go

I’ve been out of the Java scene for a really long time, but will be coming back to it soon. I’m curious - these performance issues described here, are they inherit to how Java itself? Is it baggage from Spring/Boot? Are there ways to get more bang for the buck with some careful choices in a system like this?

The closest I’ve done to Java recently is C#, which I think may have similar challenges, but overall didn’t seem quite as bad if you avoided lots of framework extras. It also wasn’t something I was digging into deeply though, so perhaps I’m mistaken.

Spring by far - Spring is effectively a build tool running at run time (scanning, enhance, code generation, etc.) - most of it is just startup as JVM does a decent job at optimizing the cruft. In most cases the boot times don't matter, though - at least for most people, esp. when it comes to production. (it's mostly developers time wasted)

I have some personal experience optimizing Spring to record previous runs and effectively cache resolution and code generation, etc. for massive boot latency improvements. Never got around contributing it back, though (not a real spring user)

You can write clean beautiful Java and make it perform beautifully if you're careful with what you pull in for dependencies... but then you're not really using Java to its full potential as an ecosystem for basically building massive line of business apps and services that you can spin up everything from junior to offshore to 30 years of experience developers. The tradeoff for this is performance and it's a thing a lot of companies are willing to trade off on.

Things get messy the moment you involve large frameworks, reflection, ten different approaches to parallelism etc. There's also the problem of code base evolution in Java. Code before JDK-8 feels different than code after it and there's a lot of Pre-JDK 8 code out there in the wild.

And that's just it at the end of the day... it's a cultural difference in the ecosystems.

My experience with very simple Jersey web apps is ~1.5 seconds to start up. Much less than his reported ~8 seconds with Spring Boot, but still not in the 100 ms range he reports with Go. I assume one second or so is about as low as you can go with a mainstream Java framework without AOT, though I'd be happy to be corrected.

> most of it is just startup

Yep, I have some Spring code in AWS ECS and it hits 100% CPU usage on start-up before dropping back to 1.5% when idling (this is with 1 vCPU I think).

But yeah I remember reading one of the Spring devs say that some (a lot?) of the runtime reflection could be done at compile time but isn't.

I worked on an app in kotlin a while back, and am currently working in a dotnet app. We can run our entire unit test suite faster than the JVM started up on that project.

Also, 8 seconds is quick in my experience for Java - I’ve seen more like 15-30

You can do AOT compilation with GraalVM to reduce both startup times and memory usage, but then you don't have the JIT.

Having done a lot of Java and Go, Go has much better mechanical sympathy between the language, libraries, and vm than Java does. The JIT GC in Java are marvels of engineering, but they have to be.

As an example, in Java, everything is a pointer, so pointer chasing all the time, which is not good for cpu cache, etc. In Go, there is first class support for composition.

The other main adjustment, if coming from Java, is reduced cognitive overhead. It usually only takes a week or two for an experienced Java dev to be reasonably effective in Go, but it takes a few months to break the mental habits of overthinking everything.

> Spring devs say that some (a lot?) of the runtime reflection

It's a lot more than reflection, if it'd have been reflection alone - it'd be markedly better. (and yes, lots and lots can be optimized). Spring effectively:

  scans the classpath for resources - that includes jars, file system
  loads every single class matching the scanned directories as a byte array
  parses it in java (not by JVM) to check what annotations it has (it doesn't load  the classes actually)
  builds dependency tree
  enhances the previously loaded byte arrays, i.e. generates different byte code
  loads the newly enhanced classes and create instances (usually through standard reflection)
  makes calls like PostInit (life cycle)
  in some cases it uses the standard java reflection to set fields/call methods; in lots of cases java reflection is generating (and loading) a new class (byte code) to carry the process

all the steps above can be recorded on run time (or be a step in the build process) and let the JVM just load the classes organically.

as for the 100%, spring initialization is mostly single threaded - so likely you dont have many cpus dedicated to the java process. (or you meant just a single core 100%)

Can confirm with Vert.x. I used to have a JRebel subscription for code hot swapping in another project but in my current Vert.x thing I simply don’t need it. It’s just fast.

Ah okay - yeah I'm not across the internals at all.

And yeah I assume it's a single core, but definitely a heavy start.

There's nothing forcing you to write EE style code in Java though, or depend on frameworks written in that style.

The JVM starts in milliseconds

Probably you're loading many thousands of classes...

I think this upcoming feature will let you have both AOT + JIT without needing GraalVM: https://wiki.openjdk.org/display/crac

> As an example, in Java, everything is a pointer, so pointer chasing all the time, which is not good for cpu cache, etc

Strictly speaking that's not true. It's everything is a pointer in theory to make it easier to reason with and JIT / JVM optimizing in the background.

There are primitive types and there are lots of tricks in the JVM e.g. escape analysis that places objects on the heap/stack etc.

It certainly tells me it does. i can hit 'run unit test', sit back in my chair and zone out for 5 seconds, then come back and read that the test took 22ms.

Yeah that’s been my experience too.

haha well your test did take 22ms, but starting the JVM and loading all the classes and all your dependency injection stuff probably took 5s. You can test it yourself with a tiny hello world, it's pretty fast to start.

Also, you don't have to continuously restart the server for every change. Can hot swap code. And for tests there are solutions to keep the jvm running between the code->test->fix cycle. But I've never felt the need for those, running a single test is often plenty fast as one doesn't need to instantiate the whole app, only the dependencies needed for the test. So plenty fast.

Thanks for bringing up some painful memories. Now I'm working in C and tests take about 100ms end-to-end (multicore, no frameworks). Sometimes I re-run them for no reason just to enjoy how fast everything is.

Despite all the advances in JIT, I've literally never seen it correctly optimize a HashMap<Integer> (happy to be proven wrong). Hopefully the renewed focus on value types can finally bring some sanity.

But that's our point - you're saying java is fast (and it is ripping fast once it gets going) and the startup is fast, unless it's not.

In C#, all struct generics are monomorphized and struct-based abstractions are zero-cost :)

In my experience, if you write Java like you'd write to, you'd probably get similar performance. Maybe AOT performs a bit better, maybe it doesn't because it disables live optimizations.

Java doesn't lend itself well to writing Go style code, though. You quickly end up writing very "Java" Java code. That has some advantages (code deduplication, less boilerplate) but also downsides (allocation overhead, performance impacts).

In terms of Java vs C#, I don't think you'll notice too many new issues. If you opt into using heavy frameworks like Spring Boot to solve annoying problems for you, you'll see RAM usage increase massively, but it'll also take care of a lot of annoying grunt work for you. Startup times overall are pretty similar. The biggest downside in my experience is the lack of nullable types (@Nullable annotations are a poor substitute) and some syntax improvements that are new to Java but have been part of stable C# for many years.

As long as you stick with small libraries and avoid writing too much ObjectInjectorFactoryProducerResolver style code, performance of the JVM is fine.

All of that goes as long as you're able to use modern toolkits. If you're stuck in the world of Enterprise Java with Oracle Java 8, you'll notice old Java's shortcomings a lot more.

Well no, startup is fast, period. You're probably just giving the class loader a ton of work on startup? I would start with checking that I think.

It could also be when he's hitting "run test" it's actually "compile and run"...

> Well no, startup is fast, period. You're probably just giving the class loader a ton of work on startup? I would start with checking that I think.

Ok, maybe I misspoke about the "JVM" startup. But the time between `mvn test` and it actually logging the first line of user code was in the region of 15 seconds. Every java project I've worked on has had startup time issues once they're bigger than a toy project.

Saying "the JVm startup is fine, it's just the class loader that's slow" reinforces the point of "it's slow to startup"

> It could also be when he's hitting "run test" it's actually "compile and run"...

This would be true in C# too, and in go. Both of those tools have much quicker compilation times IME than java, which is just a nice plus.

This makes recycling nodes, be it ECS or Kubernetes nodes, that run tons of Spring apps, a huge pain.

Well I could switch to a programming language that isn't so slowed down by ahead-of-time compilation. Maybe a JIT language?

That's mostly because nobody cares enough to optimize it, but if you split code into modules so your number of classes, including dependencies, is smaller, then your tests will start quickly. Usually it's dependencies and DI that's the cause of this I think.

Aren't Go's compile times about the same as Java's? Of course it depends on what tooling you use to package the jars.

BTW, I wonder if any test frameworks are using the hot reloading JVMs yet? Would also solve this problem.