java -cp "lib/*" -Xmx5g -Xms5g -server -XX:+UseTLAB -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:MaxTenuringThreshold=0 -XX:CMSInitiatingOccupancyFraction=60 -XX:+CMSParallelRemarkEnabled -XX:+CMSPermGenSweepingEnabled -XX:+CMSClassUnloadingEnabled
That gets cumbersome really quick.
Honestly my preferred jvm deployment is supervisord. I'd like a java only solution if one should exist that's good but these newer containers don't work for self made servers that don't run http.
Some of the instrumentation could be good, but at the end of the day I'm not sure how much it would help.
It looks pretty solid though. I'm still not totally sold on gradle as of yet mainly because of the tooling.
The native libs thing is a plus though. I'll see where I can work it in. Good work with capsule though.
Also, thanks for mentioning Packr even though noone has used it in production yet. It's only a week old.
Assuming OP is the original author: are you by any chance following JGO? I see Quasar is actually using Matthias Mann's green thread lib which i don't think was advertised anywhere outside JGO.
Like I mentioned last time I appreciate an overview of modern Java practices but boy howdy I can't discern if this is clever trolling or a cheap way to make me read Part 3.
The alternative are servlet containers, like Jetty or Tomcat. Jetty for example is very popular and really good and can be easily embedded inside an app for single JAR deployments, which is an awesome way to deploy an app btw.
Generally I think of Tomcat as being as much of an application server as something like JBoss is.
Granted, JBoss/Wildfly has a lot more enterprise-y features, but they both 'serve' web 'applications.'
In Java terminology, "application server" has actually started to mean servers capable of the full Java EE stack, which really means EJB and JMS. Jetty is not capable of those. Tomcat from what I know is not capable of those either. Both are targeting first and foremost the servlets API, which is pretty light and arguably good (well, at least since the latest Servlets 3.1, which finally adds asynchronous readings of requests).
Servlets is the piece of Java EE that I actually like - for everything else, there are third-party libraries and frameworks - though I've been working lately with Play framework, which comes with its own server and deployments on top of servlet containers is not officially supported, but that was primarily because it is a fully async framework and Servlets < 3.1 was inadequate for that, so things might change.
I also prefer embedding and the deployment of fat JARs to WARs. Makes things easier - a WAR implies that you need a management interface and a configured instance of your production server. A JAR implies that everything comes bundled in, configuration and all and you only have to copy and execute it directly, plus with embedding you have more fine-grained control. When I was using an embedded Jetty for example, a fine-tuned the shit out of its thread and connection pools, all from code.
These days CDI is the centrepiece of Java EE, along with services like Interceptors and Bean Validation. JPA, JSF and JAX-RS also play pretty important roles, which are all things you don't find by default in Tomcat.
Actually, nearly every time I see people using Tomcat they add many of the things mentioned above. You might as well start with TomEE then and work from there.
JAX-RS is the only thing OK-ish in the list you mentioned, unfortunately its design that ignored asynchronicity has been shortsighted (heard that it got fixed in 2.0, but apparently they also added junk).
There's no need to bundle everything up into an executable jar or war. It's much easier to write a Chef recipe to copy all the right files to the right places. And it's even easier to create a Docker container with the exploded war in its correct place. Both of those solutions don't rely on the clunky put-everything-inside-a-jar and the resulting need to write your own ClassLoader.
Few reasons contribute to that. One of them is exhaustive pure Java library ecosystem, which means that it's very unlikely that you'll run into intricacies of native library access on your host system. Another is excellent tooling, which means that when all your processes are Java processes, lots of operations-related tasks (process management & monitoring, log rotation, etc) get significantly easier.
More on that I wrote few weeks ago: http://www.mikhanov.com/2014/03/31/a-love-letter-to-java-363
Plus, there's the benefit of (essentially) standard builds, project layout, dependency management, and deployment strategies (jar or war). A java developer can walk into a new job and be productive right away.
Docker - "The Linux Container Engine"
Think you've got your answer there.
Sure, it's a nice benefit, but is it actually used? In 2014, how many of us are doing backend/service development in java and not deploying to Linux machines? There are certainly still backend/service java shops out there that aren't deploying to linux machines, but I wager that most are.
Regardless of those numbers though, why do the linux shops actually care about being able to deploy to non-linux machines if they never do in practice? Python can run just about anywhere these days, but python shops that deploy exclusively on Linux don't care about maintaining deployment compatibility with systems that they don't actually use.
In 2014, how many of us are doing backend/service development in java and not deploying to Linux machines?
If you are developing backend/service shit for other companies to use, then that is one thing, but I am thinking about in-house development. Most megacorps aren't software vendors; if they are developing software it is for themselves.
Regardless though, some megacorps deploying to windows doesn't explain the practices of java shops that don't deploy to windows. It seems like portability just for the sake of portability, without any real purpose but with real added hassle.
At the enterprise it is also common to develop on Windows and deploy across Windows and UNIX systems.
Finally, there are quite a few banks, finance and insurance institutions running Java on mainframes.
I am talking about one specific kind of development: Service/backend development done in-house in a corporation that is not a software vendor. If you're doing Swing/OpenGL work, you're not doing the sort of work that I am talking about.
In my experience, while this sort of development may deploy to different nix's at different companies, within any particular company it tends to always deploy on only type of system. Most "mega-corps" are not software vendors; when they develop software they are developing it for themselves. This means that they control the entire stack, making portability fairly pointless.
When development in these organizations is done in other languages without such a "portability culture", not a second of thought is given to portability. The second a service is written in Java though, everyone starts jumping through hoops for something that they will never use.
Portability is important, even when you supposedly know your deployment platform.
And the days of needing to use different development and production OSs are over...if you develop in Windows and deploy on some Unix variant, you should be using a VM.
I'm still not buying ClassLoader contortions as being in any way justified to support the single deployable artifact requirement. It's just not necessary and can lead to many subtle and hard-to-track-down issues as well as introducing unnecessary performance overhead.
What about OS/400, OS/390, MVS, etc?
Believe it or not, the world is a lot bigger than *NIX and Windows. Go track down a mid-sized manufacturing company in the midwest or the southeast, for example, and I'd almost bet you money they have apps running on an AS/400 (iSeries), or an S/38 or S/36 or something, if not a mainframe.
You're assuming you have control over what the customer's servers are.
- Create a fat jar, which necessitates bytecode transformations to prevent namespace collisions if you don't want to be sloppy. JarJar did this kinda of thing many years ago, but it seems hard to trust a process like that.
- Create what is essentially a self-extracting archive. This looks like the path Capsule took. This introduces unnecessary state and increases startup time. This is what servlet containers do, so reinventing that wheel seems like a particularly foolish decision. Not to mention that it appears that Capsule adds the additional step of pulling in all dependencies from an external server...slow deployment and an additional point of failure!
I still believe that some sort of VM image or container makes the most sense. Short of that, any system that supports sshd can be configured with Chef. The "solutions" I see coming out of the Java camp all seem like hacks by comparison.
The delay is going to be a lot more than 100ms if you use something like this when scaling elastically. In that situation, you need to go from automatically provisioned to running and accepting load in as short a timeframe as possible. Fully-baked machine images or Docker containers work for this use case, your solution doesn't.
Also, making your Maven repository a dependency of your production environment just seems like a bad idea. It creates an additional attack vector from a security standpoint and an additional component that can fail from a reliability standpoint.
> making your Maven repository a dependency of your production environment
It becomes a dependency of your deployment process, which probably already depends on a maven repo. And a maven repo can be just a NAS drive, which is often reliable enough for use in production.
2. Because everything is standardized, which is a lot easier than stitching together lots of pieces, each has to be deployed upgraded, configured, monitored and managed differently.
3. Because some Java libraries are used to do very sensitive stuff (performance wise, or correctness wise, or security wise), Java libraries, on average, tend to be of very high quality, at least relative to almost everything else (well, except maybe the Java browser plugin, but we're talking server-side here).
4. With Capsule there is no need to use a custom class loader, it's not any clunkier than Docker, and it's cross platform. It's just a cross-platform Java executable, which is stateless and requires no installation.
It basically downloads the new version, updates the database running liquibase, updates some configuration files that are on git and finally updates the application war on tomcat.
Not so bad for a single file script.
App servers have historically been hard to deploy, maintain and dev against. There has been a huge amount of progress in the last few years. Based on my experience, Java EE / Wildfly is actually easier to develop with than Spring Boot.
These servers are just unzip == fully installed things. We have them checked in and just deploy them automatically whenever there's a need for an update, pretty much as every other library out there.
I wonder are we missing something or do people just wrongly think there's "something" difficult, where there's no need at all for things to be difficult?
If you want to go really low level, a few of them even show the generated assembly code by the JIT.
As complement of it, check JIT Watch
I've never totally understood why people want to make fat jars. It seems like a process full of headaches since you can't have jars in jars. Wouldn't it be much easier to create a regular zip file with a small script to set the classpath and and run the project?
I'm not sure I understand the motivation for embedded instead of standalone servlet container. The article linked to some slides but they mostly seemed be demonstrating that you can function using an embedded container rather than providing clear benefits. Maybe it would have made more sense with the associated talk.
Can anyone provide more insight to these?
As to full blown app servers vs embedded servers, I think it's the other way around. It's the big app servers that require justification, as they are a lot more cumbersome to set up and deploy.
You wouldn't do this for a production deployment, right? Application starup that may or may not require access to the artifact repository to complete successfully. When that idea bounces around my developer neocortex, my sysadmin hindbrain starts reaching forward to strangle it.
And if you're not going to do it in production, doing it in development means having a gratuitous difference between development and production, which, again, is something i have learned to fear.
A zip with startup scripts is OK, but it requires installation.
'gradle installApp' works out of the box, and 'installs' the jars and scripts in your build directory, which is all you need to run locally. It's work of minutes to write an fpm  invocation that packages the output of installApp as an operating system package, which you can then send down the pipe towards production. This is simple, easy, standard, and fully integrated with a world of existing package management tools. Why would i use Capsule instead of doing this?
And i still don't see what advantage this has over just pushing out normal packages.
Alternatively, you can embed the dependencies, in which case it's just like a "normal package", only it doesn't require installation, and is just as easy to create (your build tool can create it). So it's a platform independent "normal package", with added features.
But then, i have spent a fair amount of managing machines, shuffling packages around apt repositories, writing Puppet code and so on. Perhaps for a developer who has not served a sentence in operations, operating system packages are a less comforting proposition.
You seem to be very keen on avoiding "installation". Could you tell us about why that is?
And could you remind me what the added features of a capsule are? Putting aside the differences in delivery mechanism, which as i've said, i'm afraid i see as misfeatures, the only one i see is that it automatically finds the right JRE.
(Sorry you're being downvoted, by the way. I think this is an interesting discussion, and your gracious responses deserve upvotes, not downvotes.)
There are other ways to achieve stateless deployment, and Capsule is not always the right fit. For example, it's not the right fit if your app requires an Oracle DB installed locally (it could be done, because Capsule supports an installation script embedded in the JAR, but that probably wouldn't be a good idea in this case). But when it is the right fit (e.g. microservices, grid workers etc.), it's much more lightweight than any other solution.
Again, for what advantage?
I mean, fine if you want to do that. But I wouldn't call it Modern Java or recommend anyone else do it.
And if you don't want to enable automatic upgrades, you still get stateless, installation-free deployment.
And you can't see why this is a bad idea?
Please don't. Never assume that production machines have access to the internet or even an internal Maven repository. If they do, they're a security risk. There are plenty of scenarios I can think of where an attacker could possibly make your application download something from the internet, or has your application server functioning as an anonymous proxy by exploiting your application.
Aren't the app servers and the embedded servers the same servers just configured differently? Standalone it is configured with XML and and embedded it is configured with Java code. It doesn't seem to me like there is much of a set up difference.
As far as I can tell, it comes down who cares more about the settings for the server, the people who maintain the machines or the people who maintain the code. If those people are the same then it sounds like it really doesn't matter.
> As far as I can tell, it comes down who cares more about the settings for the server
There's a lot of truth to that. I think that the whole devops approach is that they should be the same people. Of course, you might have valid reasons for not subscribing to that approach.
I grant you, upgrading servers can be more involved depending on how you do it.
 - http://tomcat.apache.org/tomcat-7.0-doc/deployer-howto.html#...
Fully agreed with the zip and script though, that's what we do (using the appassembler maven plugin, so it just happens as part of the build).
I'm not sure why it matters that different distros put Tomcat in different locations. Embedding Jetty just means that now you have to restart every time. You could just do the same thing with standalone Tomcat instance. The only reason you are running out PermGen space on the redeploys is because your application is not cleaning up its threads on shutdown.
Why is it difficult to do any of those development things without having the server embedded in your application? Pretty much every Java build tool (Maven, SBT, Gradle, Lein) makes those things easy to do without specifying an XML file for your server. Also, they will frequently let you reload your application, which is faster.
So now either I need to figure out how to hook my deployment script into the host system's init scripts, or I need to package tomcat as part of my app (and worry about its shared library linkage). With embedded, my app is just a program, not at all integrated with the host system; it runs like a program and stops like a program.
> Why is it difficult to do any of those development things without having the server embedded in your application?
Because the tool needs support for the server and often it has to be done in a server-specific way. How does one use JProfiler on an app deployed in Tomcat? One edits one of the internal Tomcat xml files. How does one use JProfiler on an app that embeds Jetty? One runs it with the same command line arguments one uses for running any other app with JProfiler. And it's like that for every tool - the way to do it with Tomcat is slightly different, the way to do it with an embedded server is standard.
> Pretty much every Java build tool (Maven, SBT, Gradle, Lein) makes those things easy to do without specifying an XML file for your server.
If you're building a war you need a web.xml (or I'm seeing some speculation it can be done with annotations? Either way, much less clear than just making a servlet in code).
> Also, they will frequently let you reload your application, which is faster.
Not in a way that matters. If you're just doing day-to-day dev then IDE hot code replace is faster still. If you want to be sure you've cleaned out all traces of the old version and are definitely running your latest code then reload isn't reliable enough.
Tool support is not server specific. In both cases you set the appropriate VM arguments. What XML file would you have to edit?
Yes, hot code replacement is faster (although it comes with other tradeoffs). Why can't you do that with standalone Tomcat? What is unreliable about reloading? You can leak PermGen space if you don't finalize your threads properly, but why wouldn't you be running the latest code?
You're right about the web.xml file. I had forgotten about that. I can see why you might prefer to write that in Java code rather than XML.
With an embedded Jetty my app can just be a jar or set of jars. Tomcat has a directory structure and ships as a binary rather than a jar so I'd need to unpack it and worry about it being linked correctly for the target system.
> Tool support is not server specific. In both cases you set the appropriate VM arguments. What XML file would you have to edit?
I forget the name - catalina.xml or something? You can't just run ./tomcat -agent:myagent.jar. because it won't pass the arguments through to the JVM, unless that's changed.
For setting JVM arguments, I think you are thinking of 'setenv.sh' or 'setenv.bat'. These are shell scripts where you can set environment variables used when running Tomcat. This includes the standard 'JAVA_OPTS' to set VM arguments.
I find Maven's abstractions surprisingly hard to understand. It is a major part of my annoyance with Java as a dev setup. I miss Make/Ant. Not sure if Gradle will be a good replacement or not but would be curious to try.
Part of the problem is that the initial steps are so easy that you're actually never forced to learn the mechanics of what is happening. Hence you get away with a lot until you need to do something there isn't a magic command for. This is also why you see blithe statements that you "don't need to know Groovy to use Gradle", etc. However these are only said after the fact when you've learned Gradle, understood enough Groovy to intuit what it's doing and then retrospectively realised there is a theoretical path from ignorance to enlightenment that didn't involve learning Groovy.
What you would probably like is Gant, which is Ant entirely converted to Groovy. It's missing all the higher level project stuff from Maven / Gradle, but it works a treat when you just want the simplest possible thing that could work.
> ... realised there is a theoretical path from ignorance to enlightenment that didn't involve learning Groovy.
Yes! Learning is so non-linear that sure there are in theory fine lines that minimize the amount of new stuff to learn. Yet, it never is that simple. Bad design, indeed.
> What you would probably like is Gant, which is Ant entirely converted to Groovy. It's missing all the higher level project stuff from Maven / Gradle, but it works a treat when you just want the simplest possible thing that could work.
I will check it out; thank you!
Currently I'm in a large enterprise project where the build department decided it was a good idea to reimplement a whole complex build in gradle (basically all that maven does by default). It didn't help that they didn't realize Gradle uses Maven conventions by default, ignored that and ran their own dependency system. The build is REALLY slow and really complex and nobody has any idea exactly what it does.
In ant having a non-standard project layout was ok, but then at least the build was FAST and didn't take half a minute just to START the build system.
I'm not yet sure if gradle actually IS better if you use it right, or if it's just a very very slow solution to a problem you shouldn't have it the first place if you do it right with something like maven.
My woes are mostly around Android's gradle integration, which appears poorly documented at best. There are tons of magic commands that appear out of nowhere and with no explanation beyond "use this.". And Gradle's massive guide covers things that barely see the light after being seemingly wrapped in several layers by Android's plugin.
I want to give it a real try, but the Android projects I deal with are reasonably complex, would use code generation (`scopes.PROVIDED.plus += configurations.provided
}`? wtf does that even do and how in the world would I have discovered I could use it?), and need multiple products. I would have to know what and why before considering switching over for real, but there seem to be some huge walls preventing me from finding out.
The DSL is Groovy syntax from Groovy's antiquated Antlr 2.7 grammar, so simply by using Gradle you're using Groovy along with all its warts. Underneath, Gradle isn't so much a DSL as an API shipping with a programming language. You could just as easily write the first Gradle example from the article in most other JVM languages. If it was in Clojure...
(require gradle :as g)
(g/apply :plugin "java")
(g/apply :plugin "application")
(g/jvm-args (str "-javaagent:" (-> (configurations.quasar.iterator) next))))
I took a totally baroque project setup (they just don't know WTH they're doing) with some JUL and some log4j, piped everything into SLF via the bridges, then configured a log4j backend (which I'm confortable with) and now I'm the only person in the project that actually has control over their logging and configures it to log into nicely separated files for application output, SQL statements and so on :-)
In an ideal world we'd all just use one log system, but Sun screwed that when they created the very unnecessary JUL back in the day. SLF solves the logging mess nicely.
If there is one logging engine to depend on that libs can use, this means less dependency hell.
Fighting it is a losing battle with how much traction it has among all the different libs.
Also consider the case of third-party libraries. I have dependencies in my application that are hard coded to log to about three different logging implementations. If they instead used SLF4J, the end user could choose whichever implementation they were already using in their app, instead of having multiple implementations or needing to redirect the output via an SLF4J bridging module.
Other useful tool for logging, that integrate with logback or log4j is sentry:
And to use it with java:
edit: Further to the point of getting Java server technology to a simpler to understand, design, debug and maintain place, SparkJava makes it insanely easy to do REST services eliminating both Annotations and heavy configuration. I've also speed tested my side projects on SparkJava and they are unreal fast.
The only alternatives, besides Excelsior, is the new RoboVM, Websphere Realtime JVM, Aicas, Aonix among a few other ones.
And if Graal eventually replaces Hotspot on the reference JVM, it might be that Truffle comes along as well.
Heck, the runtimes you're advocating are generally installed as binaries. If they can get updated, then so can any other binary.
And if a user doesn't bother to update a given application's binary when a critical flaw of some sort is found, then there's a very good chance they wouldn't bother to update any runtimes that are installed, either. This is true even when some sort of update notification and installation process is offered. The end result is the same in either case: the update is not applied.
Runtimes shouldn't be portrayed as any better in this case, when they're generally no different than any other binary.
Of course, your project probably includes its own logback.xml in your resources directory.
Libraries are typically distributed as JAR files, so when you include their JAR on your classpath you're also including their logback.xml in that JAR.
Now, when you go to package up an uber-JAR for your project, whatever packaging tool you use has to merge all those files into one classpath. So now it's found collisions: there are multiple, different logback.xml files on your classpath but you can only have one. How does the tool pick?
So now you have to specify a "merge strategy" to make sure that only your logback.xml is merged in and the rest are discarded.
logback.xml is just one example. Typically Java classes themselves merge just fine because the packaging tool can dedup by fully-qualified class name which is almost never shared across projects. So even if two different dependencies have a Utils class, the FQCN is different. com.foo.bar.Utils is not the same as com.bar.baz.Utils, so there's no conflict.
Conflicts can also arise from libraries which include different versions of the same transitive dependency. Generally the fix here is to exclude a transitive dependency from one of your direct dependencies so as to only include one version of that library. A common culprit is Apache Commons since so many Java libraries rely on it.
Only then JBoss decided that all their stuff (including Hibernate) needed a facade of its own, which manages to not quite properly interact with SLF4J!
It's almost enough to drive me to printlns.
All of a sudden all of the log4j.xml files that everyone has sprinkled everywhere end up overwriting each other...
Since each JAR is just a zip if you have two files in seperate JARs with the same path you'll have a problem when you try to merge them.
I do think a well-rounded Java developer should know about OSGI (i have to confess that i don't). But it would be mistaken to think that it's mainstream.
I think it's one of those ideas that sounds great in theory, but in reality it all falls apart.
Even if it worked as advertised, fact is many 3rd party libraries have all sorts of issues with OSGi. For example, my company is stuck on Jersey 1.x because 2.x doesn't work properly with OSGi.
Testing is much slower and harder to write. Testing seems to be an afterthought....
I think you can get most of the benefits of OSGi simply by using proper dependency management (Gradle). Use the Single Responsibility Principle and IoC and you'll get good modular code. These are much easer to do, have low risk, and are easy to test.
After all, you want to profile your app's hot spots ;-)