If we're going to set up JakartaEE to be a platform for the next 20 years then we're going to have to do some wildly different things alongside supporting what got us here in the first place. Profiles may be a way to help with that but not (in my opinion) if
we restrict the way we think about them as being simple concentric circles around a single core. Different style of programming are going to need different profiles (defined sets) of technologies that may share some in common without necessarily having to
have a single common minimum set of core technologies in all profiles. Reactive streams are a good example. The servlet spec (for example) is important to JakartaEE full and web profiles but doesn't need to be part of some future JakartaEE ReactiveStreams
profile (if we worked on such a thing). It would still be part of the profiles for which its relevant including the full profile.
I see near universal agreement that the the JakartaEE full profile will remain important, and it will grow. If we get new and smaller profiles right then they can support platform innovation without having to drag everything forward at the pace of the oldest.
So I agree with James' final sentiment by thinking about it without the 4-letter c word.
> I think a cloud native Jakarta EE is one that treats servlets as an option chosen for the right use cases, not as a default in the core.
Regards,
Ian
Ian Robinson
Date: 22/05/2018 03:54
Subject: Re: [jakarta.ee-community] About Profiles / Servlet API
Not sure where to put this reply since the thread seems to have split a couple of times, but to clarify my original suggestion.
When I say synchronous communication, I'm not referring to IO. Whether the servlet API is blocking or not is not a consideration to me at least in whether it should be part of the core. Rather, the question is whether synchronous communication - ie, request/response,
where both parties in the communication need to be actively participating in the communication for it to succeed, where one party being down leads to the other unable to proceed, synchronous communication is what I don't think should be in the core. The servlet
API is inherently synchronous, if a client tries to send a message to a servlet, and the servlet isn't running, is overloaded, or is failing, the servlet won't get that message later when the problem is fixed, this is because it is based on synchronous (dictionary
definition "at the same time", ie, both parties active at the same time) communication, whether it's using non blocking IO or not.
The alternative, asynchronous communication, which typically uses message brokers as an intermediary transport, does not require both parties to be participating at the same time, so the sender can send the message regardless of whether the receiver is able
to receive it at that time. In a world where deployments now consist of tens, hundreds even thousands of services, requiring all services to be running at the same time, which synchronous communication does, leads to a system that is fragile and far worse
than a monolith. Hence, asynchronous communication is imperative in this world of cloud native services.
I'm quite aware that Servlets aren't tied to HTTP, but they are tied to request/response, and this is the issue.
I anticipate that in 5 years, most backend services will primarily just use asynchronous communication - the primary places that will use synchronous communication will be services that directly face users, where users need a timely response to their requests.
In many architectures today, this is already the case.
That's not to say Servlets are ever going to go away - it's not an either or, both synchronous and asynchronous communication are needed in all systems. But the current situation, where synchronous is the primary means of communication, reflected by the fact
that the servlet spec is in the core, is coming to an end, and it's cloud native that is signalling the end of that. I think a cloud native Jakarta EE is one that treats servlets as an option chosen for the right use cases, not as a default in the core.
You are aware that each of these standards are defined elsewhere (e.g. OASIS) therefore it is important to separate the specs (if you propose something underneath MQTT or AMQP this is simply the wrong place) and implementing solutions (this is what can be shaped
or influenced here)
Werner
JMS is for applications accessing MOMs, which in turn use AMQP/MMQT.
What I proposed is a layer UNDERNEATH AMQP/MMQT.
-Markus
Sent: Montag, 21. Mai 2018 19:47
To: Jakarta EE community discussions
Subject: Re: [jakarta.ee-community] About Profiles / Servlet API
Of course AMQP / MMQT are message-based, so they are by nature closer to e.g. JMS.
I can't say, what the just forming EE4J Messaging/JMS project might like to do about this, but having worked in a cross-cut between MQTT and JMS for major clients, I see potential for synergies between EE4J and IoT especially now they
are both under the Eclipse umbrella
Werner
I also thought so but in fact Servlet API only works for request-response communication models, which is not given for at least UDP, AMQP and MMQT. IMHO you cannot implement AMQP / MMQT ontop of Servlet API.
-Markus
Sent: Montag, 21. Mai 2018 18:23
To: Jakarta EE community discussions
Subject: Re: [jakarta.ee-community] About Profiles / Servlet API
I’m not up to speed on everything Servlet, but I do like the idea of breaking it up so that you have an HTTP specific Servlet and a more general Servlet programming model that can be extended to the semantics of other application protocols. I think, to some
degree, this already exists its just a matter of the Servlet provider supporting the stack desired (e.g. UDP/TCP/SCTP) and then having a higher level programming model for each application protocol (e.g. HTTP, AMQP).
I'd like to extend this idea by a new API for the UDP/TCP/SCTP layer: It might be useful to build the HTTP API layer (= former Servlet API) ontop of a new UDP/TCP/SCTP API layer which allows Servlet engines to replace the underlying transport
technology easily. That API could wrap micro-servers dealing solely with data packets instead of application protocols. So not only HTTP would run ontop of that, but also AMQP or MQTT or even FTP. The bean driven by such a system would be a "Stream-let" or
"Protocol-let" for example. So re-implementing the Servlet API would mean to write a "HttpProtocolLet", AMQP would use an "AmqpProtocolLet" etc. In the IoT-Environment this could be beneficial, and it would simply re-use the technical platform already existing
e. g. in Grizzly.
-Markus
Sent: Montag, 21. Mai 2018 17:58
To: Jakarta EE community discussions
Subject: Re: [jakarta.ee-community] About Profiles
Hi,
I think that James didn't mean to toss out HTTP processing but that the servlet API isn't well suited for reactive processing. The Servlet spec is also huge, comparable to EJB and could be designed and split into multiple specs. Or even obsoleted by another
more modern spec in a similar way as CDI is now obsoleting much of the EJB spec.
I don't quite see the Servlet spec being deprecated / obsoleted, like EJB was.
I do like the suggestion of splitting out the Servlet spec in parts. Specifically we could potentially have a core HTTP engine (like Netty, Grizzly, and Coyote), a "bare" Servlet API layer, and something like a higher level CDI based layer (making a simpler
Servlet that's a proper CDI bean).
Kind regards,
Arjan
Cheers,
Ondrej Mihályi
Senior Payara Service Engineer
Payara Server – Robust. Reliable. Supported.
----------------------------------------------------------------------------------------------------------------------
Payara Services Limited, Registered office: Unit 11, Malvern Hills Science Park, Geraldine Road, Malvern, WR14 3SZ
Sent: 21 May 2018 16:03:00
To: Jakarta EE community discussions
Subject: Re: [jakarta.ee-community] About Profiles
James,
I think you have missed something in the Servlet programming model. While its primarily used for HTTP requests its not bound to that model or even synchronous request/response messaging. Reactive Streams could work well within the Servlet model or form another
programming model. If I’m wrong than Serlvets have changed more than I thought since I used them for pure TCP/IP processing back in 1998.
I’m an advocate of Reactive Streams as a new programming/processing/messaging model, but it would be premature to toss out HTTP processing and Servlets in order to embrace Reactive Streams as the “one true” model. Having been around for a long time I can tell
you that today’s silver bullet often becomes tomorrows legacy nightmare. I’ve lived through that transition at least a dozen times. There is no “one true” model. There is only those options that most useful now.
I don’t believe that is the fate of Reactive Streams to become legacy, but at this point its not a widely adopted programming model and I think we need to serve the needs of the mass majority of developers who use HTTP and Servlets IN ADDITION TO introducing
and supporting new methods like Reactive Streams.
I would not make Reactive Streams a required part of the Core at this point, but as Reza pointed out the Core can evolve over time just as well as any profile. If Reactive Streams is introduced as a Specification and wrapped into a Profile on top of the Core,
it could remain very lightweight while providing a powerful programming model for those interested in that paradigm.
Are servlets really necessary in the core? Yes, they may have been central to Java EE for as long as Java EE has existed, but things are changing, systems can no longer be seen is a big static state store that can just be queried and updated with synchronous
communication, rather they are being build using streams, where the current state is in a constant state of converging, but never actually getting there, and communication is primarily asynchronous. Look at things like Kafka Streams and AWS Lambda and Azure
Event Grid - event based systems that are only concerned with asynchronous messaging are rising rapidly in popularity at the moment. And this isn't even that new, almost 10 years ago Heroku had both web dynos and worker dynos - worker dynos had no HTTP interface,
and you could argue that deploying something that started an HTTP server to a worker dyno was overkill. Now is a perfect opportunity to realign Jakarta EE with current industry trends.
And even for technologies that use synchronous communication, look at the rise of things like gRPC - this does use HTTP, but not on top of servlets. No one wants to deploy both a servlet HTTP server and a gRPC server, that's too heavyweight. Other things like
gRPC may well surface, do we want the servlet API to get in the way of people using these new technologies with Java EE?
As a counter point against requiring servlets, the MicroProfile messaging spec currently being developed will have no dependence on servlets, and I anticipate that there will be many use cases where you'll deploy services that use nothing but MicroProfile messaging
for communication, plus a database.
Perhaps as little as 2 years ago, I would have agreed, servlets are core. But I think there's a big shift at the moment, and a decision to make servlets core today could leave Jakarta EE behind.
You know I have a tremendous amount of respect for you, but as I suggested the truth is that none of this is about technical merit. Sometimes we really do need to just bow to where the market is driving us and not fight it. Once we have the technology on firmer
footing is the time to explain to the marker why we are right. That's not now.
The market demand from where I clearly have seen again and again is allowing people to start with Java EE with nothing more than Servlet and a la carte let them add whatever else on top. In addition there is a viable smaller market for one or two sensible profiles.
Right now people also want fat jars and hollow uber jars.
Giving people basically what we've been trying to push for years plus some ability to remove things or define things is yet another road to fighting another uphill battle that's honestly tiresome. The end result of where we are today should be telling us loud
and clear we need to be thinking about these things differently going forward.
Sent via the Samsung Galaxy S7, an AT&T 4G LTE smartphone
-------- Original message --------
Date: 5/20/18 4:40 PM (GMT+01:00)
Subject: Re: [jakarta.ee-community] About Profiles
Hi,
I am totally with Mark on these particular issues. Whatever the actual merits of the argument that we can argue endlessly, the reality is that customers cite Java EE "weight" as a reason to not adopt it with Docker/Cloud/Microservices, etc.
It's true that this is being cited, although it's indeed "weight' between quotes. With only a few exceptions I've seen the actual weight (startup time / memory size / footprint on disk) of the supposedly lighter weight solution actually been approximately the
same size or even bigger.
People arguing that they use e.g. Tomcat because it starts in 500ms and is only 7MB in size, but then they add 50MB worth of libraries to /lib and an other 40MB of libraries to WEB-INF/lib. The combined whole being just under 100MB in disk space.
As a second point, a (standard) static config + pruning tool would again address this concern, and I argue it can do even better than any profile ever can.
Kind regards,
Arjan
Additionally I believe other than minimizing endless conflict with Spring folks over CDI, making clear what Java EE is and how well adopted/not it is, a Servlet only core is also the right answer to combat the criticism that Java EE is fat.
Sent via the Samsung Galaxy S7, an AT&T 4G LTE smartphone
-------- Original message --------
Date: 5/18/18 3:47 PM (GMT+01:00)
Subject: Re: [jakarta.ee-community] About Profiles
Hi,
This isn’t just about vendor choice. You are certainly not alone in being happy with the full profile option. However, there are other classes of users/developers that aren’t and these have existed since the dawn of J2EE. For example, some people want to deploy
their favour app server on to constrained devices which may be running on the cloud where an additional 50MB costs real money when run for hours or days or weeks or longer.
I wonder, is there in practice any service / device where a mere 50MB of disk space makes all the difference?
You mean apart from cloud (yes, those 1 cent costs do add up, and there’s private cloud deployments of which you may be unaware and have some funky architectural choices behind them), IoT? Oh and maybe it’s not just 50MB but some modern implementations can
still be a bit “plump” in some areas ;)
Some developers want to reduce the maintenance complexity or boot time of their favourite app server by stripping out those capabilities they don’t want.