[undertow-dev] scala, activate, nio, async jdbc, future, threads

[Stuart Douglas]

You need to call the HttpServerExchange.dispatch() method, which will prevent the exchange from being ended once the call stack returns. 

There is however a thread safety issue here, as you can have your callback being called at the same time as the call stack is still returning, which means it is possible to have two threads using the HttpServerExchange, which is not thread safe (in practice this will *probably* work fine, depending on how your code is written). 

The correct way to deal with this is: 

HttpServerExchange.dispatch(SameThreadExecutor.INSTANCE, new Runnable() { 
public void run() { 
//setup callback 
} 
} 

Basically this does two things, calling dispatch means that the exchange will not be ended when the call stack returns, and dispatching using SameThreadExecutor means that the runnable will be run immediately after the call stack returns in the IO thread (or whatever thread is currently active). 

If you don't want to use a callback and just want to use await: 

HttpServerExchange.dispatch(new Runnable() { 
public void run() { 
//call await() 
} 
} 

This will dispatch to the worker thread pool, and await() will be called in a worker thread, although if you use this approach you might as well just use plain JDBC, as you loose the benefits of non-blocking IO. 

Stuart 

----- Original Message -----

> From: "Laurent Bedubourg" <laurent at labe.me>
> To: "Jason Greene" <jason.greene at redhat.com>
> Cc: undertow-dev at lists.jboss.org
> Sent: Saturday, 13 December, 2014 6:58:51 PM
> Subject: Re: [undertow-dev] scala, activate, nio, async jdbc, future, threads

> Thank you very much Jason, it helps me a lot.

> One last remark regarding " you can save yourself a worker thread and reuse
> the thread they generate the callback from. At that callback point you can
> write your response out to the exchange."

> When I try to do this, calling x.getResponseSender().send(str) un the
> onSuccess callback of the database future, I got an exception
> java.lang.IllegalStateException: UT000004: getResponseChannel() has already
> been called.

> I imagine that the response has already been somehow cleaned up / modified /
> terminated by Undertow when the handler exited (I use SessionAttachHandler,
> date() and routing() in my test).

> I could not find a way to tell Undertow that I would handle the exchange and
> end it manually.

> I am using undertow 1.1.1-final maybe this has been fixed since this version?

> Thanks again

> Laurent

> On Sat Dec 13 2014 at 4:57:33 AM Jason Greene < jason.greene at redhat.com >
> wrote:

> > Hello Laurent,
> 

> > Your hunch is correct. You never want to block the I/O thread as each
> > thread
> > is shared between many connections, and a pause prevents them from being
> > processed. It’s totally reasonable to call dispatch(), which transfers the
> > exchange to the worker pool where you can wait on a database reply, and in
> > fact this is a normal pattern. The worker pool is sized to be pretty large
> > as its expected to host lots of waiters. You can of course tweak this, or
> > dispatch to your own special pool if you like. The big advantage to an
> > async
> > database call, even with requiring a worker thread, is that you can do
> > other
> > work while its running if your pattern fits (e.g. parallel db calls,
> > iterative computation, etc).
> 

> > Alternatively, if the database provides a callback on complete option, you
> > can save yourself a worker thread and reuse the thread they generate the
> > callback from. At that callback point you can write your response out to
> > the
> > exchange. Just note, that the thread policy of an exchange is that only one
> > thread at a time can interact with it. If you have a use case where the a
> > request bounces back and forth (e.g. a very long stream containing multiple
> > records that must be written to the database as they com in), then you will
> > need to coordinate locking around that. Such use-cases though should be
> > weighed against the simplicity of just blocking in a worker thread.
> 

> > To answer a second question you have in a comment from your example,
> > startBlocking() just enables the use of a plain inputstream/outputstream,
> > which can only be done once you have dispatched and blocking i/o is now ok.
> 

> > Hope this helps,
> 

> > -Jason
> 

> > > On Dec 12, 2014, at 3:44 PM, Laurent Bedubourg < laurent at labe.me > wrote:
> 
> > >
> 
> > > Ok, I think my mail could be reduced to :
> 
> > >
> 
> > > Is it possible possible to transfer the HttpServerExchange responsibility
> > > to another thread ?
> 
> > >
> 
> > > Regards
> 
> > > Laurent
> 
> > >
> 
> > >
> 
> > > On Fri Dec 12 2014 at 7:11:01 PM Laurent Bedubourg < laurent at labe.me >
> > > wrote:
> 
> > > Hello,
> 
> > >
> 
> > > I am evaluating Undertow and trying to fit an async database (with scala,
> > > Activate and https://github.com/mauricio/ postgresql-async which use
> > > netty
> > > and nio).
> 
> > >
> 
> > > The good news is that it more or less works.
> 
> > >
> 
> > > My main concern is that the database system create threads using scala
> > > futures and that I am forced to "Await" for the result of my Futures to
> > > send the result to the HttpServerExchange.
> 
> > >
> 
> > > I am forced to dispath() the request because Await locks the current
> > > thread
> > > and I shouldn't lock the IO thread from my Handler, should I?
> 
> > >
> 
> > > It made me wondering : is is really a good idea to use futures + Await
> > > with
> > > undertow and more generaly nio?
> 
> > >
> 
> > > If the thread pool is limited and I lock threads, even if the database
> > > driver is async and uses nio, am I bitting my own leg?
> 
> > >
> 
> > > Thanks for any input you can give me and sorry if it's too scala related
> > > :)
> 
> > >
> 
> > > Regards
> 
> > > Laurent
> 
> > >
> 
> > > PS: Here's the handleRequest I am using to test this, any comment or help
> > > welcome too since I am discovering the API.
> 
> > >
> 
> > > def handleRequest(x:io.undertow. server.HttpServerExchange){
> 
> > > if (x.isInIoThread()){
> 
> > > x.dispatch(this)
> 
> > > return
> 
> > > }
> 
> > > // useful or not? no change
> 
> > > // x.startBlocking()
> 
> > >
> 
> > > // Ok, we are in worker thread we can work a little with database
> 
> > > // val f = asyncTransactionalChain { implicit context =>
> 
> > > // for (
> 
> > > // a <- asyncTransactional { new AMember("aaa") };
> 
> > > // b <- asyncById[AMember]( a.id )
> 
> > > // ) yield b
> 
> > > // }
> 
> > >
> 
> > > // Simulate database work with just a future
> 
> > > val f = future {
> 
> > > Some(1)
> 
> > > }(scala.concurrent. ExecutionContext.Implicits. global)
> 
> > >
> 
> > > // send string to xchange
> 
> > > def sendReply(str:String){
> 
> > > x.getResponseHeaders().put(io. undertow.util.Headers.CONTENT_ TYPE,
> > > "text/plain")
> 
> > > x.getResponseSender().send( str)
> 
> > > // x.endExchange()
> 
> > > }
> 
> > >
> 
> > > // // this fails
> 
> > > // f.onSuccess {
> 
> > > // case Some(result) => sendReply(result.toString)
> 
> > > // }(scala.concurrent. ExecutionContext.Implicits. global)
> 
> > >
> 
> > > // // this works
> 
> > > val result = Await.result(f, Duration(1000, MILLISECONDS))
> 
> > > sendReply(result.toString)
> 
> > > }
> 
> > >
> 
> > > ______________________________ _________________
> 
> > > undertow-dev mailing list
> 
> > > undertow-dev at lists.jboss.org
> 
> > > https://lists.jboss.org/ mailman/listinfo/undertow-dev
> 

> > --
> 
> > Jason T. Greene
> 
> > WildFly Lead / JBoss EAP Platform Architect
> 
> > JBoss, a division of Red Hat
> 

> > ______________________________ _________________
> 
> > undertow-dev mailing list
> 
> > undertow-dev at lists.jboss.org
> 
> > https://lists.jboss.org/ mailman/listinfo/undertow-dev
> 
> _______________________________________________
> undertow-dev mailing list
> undertow-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/undertow-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/undertow-dev/attachments/20141214/722996b8/attachment.html