hc-httpclient-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pellerin, Clement" <Clement_Pelle...@ibi.com>
Subject RE: Controlling releaseConnection
Date Wed, 19 Oct 2016 14:14:49 GMT
I agree with you this feature is strange and need not be added to the library.
As I said before, I need to preserve this feature for backwards compatibility.
It has something to do with delaying the emission of an MDN in an AS2 gateway
so it represents the status of the receiver instead of the man in the middle.
Again, please don't ask me to change the architecture of the solution.
We don't own the application and we can't change it. Nevertheless it must
continue to work after we port to HttpClient 4.5.2

As for the work-around, I have something working but it duplicates too much
code for my taste. I think this is a nice opportunity for HttpClient
to learn what it is like to try to subclass its engine.

I am subclassing HttpClientBuilder to override createMainExec()
to return my own implementation of the ClientExecChain.
This is nice.

Unfortunately, our subclass of ClientExecChain must be in the
org.apache.http.impl.execchain package because support classes
in that package are all package private. This is annoying.

I modified the problematic code in ClientExecChain.execute()
but that was not enough. I need to trap the EofSensorWatcher
of the response HttpEntity. That requires duplicating
HttpResponseProxy and ResponseEntityProxy.
Again this must be done in the apache package.

The release should go through the ConnectionHolder since that's
where the "AtomicBoolean released" field is stored.
I thought of providing my own implementation of ConnectionHolder
with an empty releaseConnection() method, but that does not work
when ResponseEntityProxy calls cleanup(), this becomes
releaseConnection(false) which closes the connection completely.

I would appreciate if you have better suggestions for the work-around.
It would be nice to make subclassing easier too.

By the way, should ResponseEntityProxy.enchance()
be called enhance() instead?
Compare this to RequestEntityProxy.enhance()

-----Original Message-----
From: Oleg Kalnichevski [mailto:olegk@apache.org] 
Sent: Tuesday, October 18, 2016 10:48 AM
To: HttpClient User Discussion
Subject: Re: Controlling releaseConnection

On Tue, 2016-10-18 at 10:12 -0400, Pellerin, Clement wrote:
> I agree that code is correct. I never said there was a bug.
> I am asking how to override the behavior.
> We are porting our product from HttpClient 4.1 to 4.5.2 and we need to 
> preserve that feature because it is used by our customers.

Clement

I'll be happy to suggest a work-around but I am still unable to understand what it is exactly
you are trying to achieve by keeping a connection for messages without a message body, which
is quite uncommon (status 204, status 304, response to HEAD, and that is it). 

What is exactly the point of doing so in th first place?

Oleg 


> It would be a shame to duplicate all of MainClientExec because of a 
> lack of modularity in that class.
> 
> We tried to subclass the PoolingHttpClientConnectionManager
> but that does not work because the connection state is actually stored 
> in the ConnectionHolder created directly in MainClientExec.execute().
>         final ConnectionHolder connHolder = new 
> ConnectionHolder(this.log, this.connManager, managedConn);
> 
> Basically, I'm asking if the maintainers could consider a method like
> 	protected boolean isEarlyRelease(HttpEntity entity) {
> 		return entity == null || !entity.isStreaming();
> 	}
> 
> Or a release strategy interface if this is a common requirement.
> I am also interested in any other work-around.
> 
> -----Original Message-----
> From: Shawn Heisey [mailto:apache@elyograg.org]
> Sent: Monday, October 17, 2016 9:24 PM
> To: HttpClient User Discussion
> Subject: Re: Controlling releaseConnection
> 
> On 10/17/2016 3:22 PM, Pellerin, Clement wrote:
> > Our customer needs to delay the release of the connection until the response is
fully processed.
> > They want to turn off the early automatic release of the connection and do it manually
later.
> >
> > This is the problematic code in MainClientExec
> >             // check for entity, release connection if possible
> >             final HttpEntity entity = response.getEntity();
> >             if (entity == null || !entity.isStreaming()) {
> >                 // connection not needed and (assumed to be) in re-usable state
> >                 connHolder.releaseConnection();
> >                 return new HttpResponseProxy(response, null);
> >             } else {
> >                 return new HttpResponseProxy(response, connHolder);
> >             }
> 
> Mostly an end-user here, with no status to speak of in this project.  I do have status
on another Apache project that utilizes HttpClient, but I don't know much about that part
of the code.  I have written some HttpClient code for a completely unrelated project of my
own, but that code is VERY simple.
> 
> When I read the code above, what I see is this: It only releases the connection if the
entity is nonexistent (null) or the entity is NOT a type that uses streaming.
> 
> I will fully admit that my experience with HttpClient is limited, but I think the chance
is very small that the HttpComponents committers have made a mistake here.  I think this particular
code has probably been discussed and examined, then ultimately validated as correct.  Here's
why I think they didn't make a mistake:
> 
> If the entity object is null, then the response probably doesn't HAVE an entity (response
body), so it will be entirely self-contained, consisting of headers only, and the connection
doesn't have anything further to send.  If the entity exists but doesn't utilize streaming,
then I think it's likely that the entity was received in its entirety and has been incorporated
into the response object already, and once again, the connection isn't needed.  If my limited
understanding of non-streaming entities is correct, they have the potential to be very dangerous
from a memory consumption perspective, and my own usage of HttpClient (where I did not set
anything related to the entity type) suggests that streaming entities are used by default.
> 
> Restating in another way:  In the first situation that results in a released connection,
there's nothing to consume, you just need the response object that you already have.  In the
second situation, the entity you will consume is probably already available within the response
object and doesn't need the connection.  The comment on the release call in the code quoted
above implies that this is how things work.
> 
> In these situations, why do you need the connection to stick around?  I think it can't
do anything else that's useful for that request.  I would imagine that if the connection utilizes
keepalive/pipelining, that it will typically remain open after release and can be utilized
again for a different request.
> 
> Someone with more direct knowledge of HttpClient's internal implementation will need
to confirm whether or not I'm correct in what I've written.  My understanding could be wrong.
> 
> Thanks,
> Shawn


---------------------------------------------------------------------
To unsubscribe, e-mail: httpclient-users-unsubscribe@hc.apache.org
For additional commands, e-mail: httpclient-users-help@hc.apache.org

Mime
View raw message