[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
[paho-dev] R: R: R: Paho Java Client retransmission implementation

Hi,
It was a bug in my code. The test now succeeds. Thanks for you support.

Ciao,
 Cristiano 
________________________________________
Da: paho-dev-bounces@xxxxxxxxxxx [paho-dev-bounces@xxxxxxxxxxx] per conto di Nicholas O'Leary [nick.oleary@xxxxxxxxx]
Inviato: mercoledì 8 agosto 2012 20.07
A: General development discussions for paho project
Oggetto: Re: [paho-dev] R: R: Paho Java Client retransmission implementation

Hi Christiano,

the scenario you describe is a pretty basic one that is well covered
by routine testing the client gets hit with - although it is always
possible you're hitting a edge case somewhere.

> Most of the time there are many messages in-flight (10-15 max).

The paho client allows a maximum of 10 in-flight messages at any one time.

> At the end I sleep 10 seconds before disconnecting and I check if the
> tracking list is empty.
> It can happen it's not empty. Thus some messages are not acknowledged.

Does your test include a separate subscriber so you can confirm what
messages are actually delivered?

If you don't timeout after 10 seconds, but leave it running for longer
do the messages eventually get acknowledged? It is possible your 10
second timeout is not giving the client enough time to deliver all
1000 messages given the 10 in-flight limit.

Regards,
Nick

On 8 August 2012 17:29, De Alti, Cristiano
<Cristiano.DeAlti@xxxxxxxxxxxx> wrote:
>
> Hi Dave,
> I've slightly modified Sample.java in order to publish 1000 messages with
> QoS = 2 on my Mosquitto broker at localhost (on Linux). There aren't any
> other connections to the broker.
> I'm publishing at the fastest possible rate without waiting for the
> completion. I'm tracking the delivery tokens in a list and delete them when
> the confirm token comes in asynchronously in the callback. Most of the time
> there are many messages in-flight (10-15 max).
> At the end I sleep 10 seconds before disconnecting and I check if the
> tracking list is empty.
> It can happen it's not empty. Thus some messages are not acknowledged (I
> still don't know if Paho or Mosquitto drops messages or it's a bug with how
> I track messages). I cannot imagine a more reliable connection since both
> the broker and the client are running on localhost.
>
> I'm sure that TCP/IP implementations are reliable enough but brokers may
> be not.
>
> I think I'll try to add the republish method to the Paho API in order to
> allow for republishing/resubscribing at the application layer. If it works I
> will submit the patches to you.
>
> Ciao,
>  Cristiano
> ________________________________
> Da: paho-dev-bounces@xxxxxxxxxxx [paho-dev-bounces@xxxxxxxxxxx] per conto
> di Dave Locke [locke@xxxxxxxxxx]
> Inviato: mercoledì 8 agosto 2012 17.44
>
> A: General development discussions for paho project
> Oggetto: Re: [paho-dev] R: Paho Java Client retransmission implementation
>
> >> Do you think it would be worthwhile to add this feature to Paho?
>
> >We quite deliberately removed message retry from the client side, but
> >I don't recall the full reasoning (this was >2 years ago now). That
> >said, it does presume you are running against a 'reliable' server,
> >which, in the context of WebSphere MQ, you are and that you have a
> >reliable TCP connection (reliable in the sense packets don't
> >mysteriously go missing even if the connection is fragile).
>
> >So, perhaps there is a place in for this behaviour in the client -
> >Dave/Ian, can you recall why we did this?
>
> One of the original MQTT Java clients had retry capability built in. At
> the time it was added to handle badly behaving TCPIP implementations on some
> "wireless" networks (in the late 90s) . For instance a QOS 1 publish would
> be sent but puback on occasions would never be received. This was not the
> fault of the server but the fault of the network which never delivered the
> publish to the server.   As a result of improvements in TCPIP
> implementations the retry logic was taken out of the Java client.
>
> On the assertion thatf TCPIP behaves correctly, iIf a problem occurs then
> either the "network" informs the MQTT client or the MQTT client detects the
> problem via the keepalive. When a problem occurs the MQTT client will tidy
> up and alert the app the connection has been lost. When the app causes the
> MQTT client to reconnect (after an abnormal disconnect or normal disconnect)
> it will ensure messages that are still in-flight are delivered to the
> requested QOS.
>
> Given the background are there still reasons why retry might be useful?
>
>
> All the best
> Dave Locke
>
> Senior Inventor, Pervasive and Advanced Messaging Technologies
>
> locke@xxxxxxxxxx
> Dave Locke/UK/IBM@ibmgb
> 7-246165 (int) +44 1962816165 (ext)
> 37274133 (mobex) +44 7764132584 (ext)
> Fringe Bluepages with fuel injection
> My Cattail: Share files in IBM and save your in-box
>
>
>
>
>
>
> From:        "Nicholas O'Leary" <nick.oleary@xxxxxxxxx>
> To:        General development discussions for paho project
> <paho-dev@xxxxxxxxxxx>,
> Date:        07/08/2012 15:08
> Subject:        Re: [paho-dev] R: Paho Java Client retransmission
> implementation
> Sent by:        paho-dev-bounces@xxxxxxxxxxx
>
> ________________________________
>
>
>
> > Would I wait forever? Unacknowledged messages will stay forever in
> > internal in-memory queues and persistence?
>
> With that code, yes you would wait forever. There is also
> token.waitForCompletion(timeout) if you want to block for a certain
> period before moving on, however that still leave the message in an
> indeterminate state and unable to be resent.
>
> > I might want to track the in-flight messages in the application and
> > retransmit them if they are not acknowledged by a suitable timeout.
> > The spec allows for this by republishing the message using the original
> > message ID and setting the DUP flag in the MQTT PUBLISH header.
> > However this is not possible in Paho with the public API used in
> > Sample.java.
>
> Correct, the api does not expose that sort of republishing capability.
>
> > Do you think it would be worthwhile to add this feature to Paho?
>
> We quite deliberately removed message retry from the client side, but
> I don't recall the full reasoning (this was >2 years ago now). That
> said, it does presume you are running against a 'reliable' server,
> which, in the context of WebSphere MQ, you are and that you have a
> reliable TCP connection (reliable in the sense packets don't
> mysteriously go missing even if the connection is fragile).
>
> So, perhaps there is a place in for this behaviour in the client -
> Dave/Ian, can you recall why we did this?
>
>
> > What would be the required effort?
>
> Obviously it depends on the approach taken - here are a couple options
> off the top of my head:
>
> 1. it should be fairly straight forward to add some retry logic into
> ClientState, along with a config option for a retry-timeout in
> ConnectOptions. The client would then do the retrying under the
> covers; the application would not need to take any further action.
>
> 2. Alternatively, the DeliveryToken object could have a method
> .resend() added to it, that would resubmit the message for delivery
> (with the duplicate flag set). So, if
> deliveryToken.waitForCompletion(timeout) timed out, the application
> could choose whether to resend the message.
>
> Option 1 means retries will just work, Option 2 gives the application
> more control, but also more responsibility.
>
>
> We would also have to consider the impact on the other clients in
> Paho; we want to keep them feature compatible.
>
> Regards,
> Nick
>
>
>
> On 7 August 2012 09:36, De Alti, Cristiano
> <Cristiano.DeAlti@xxxxxxxxxxxx> wrote:
> > Hi Nick,
> > Thanks for the quick confirm.
> > I know that the spec does not require it.
> > More questions:
> > 1) What happens if we publish a message with QoS>0 and wait for a
> > confirm that never comes in (taken from Sample.java):
> >
> >     public void publish(String topicName, int qos, byte[] payload)
> > throws MqttException {
> >
> >         // Connect to the server
> >         client.connect();
> >         log("Connected to "+brokerUrl);
> >
> >         // Get an instance of the topic
> >         MqttTopic topic = client.getTopic(topicName);
> >
> >                 MqttMessage message = new MqttMessage(payload);
> >         message.setQos(qos);
> >
> >         // Publish the message
> >         log("Publishing at: "+System.currentTimeMillis()+ " to topic
> > \""+topicName+"\" qos "+qos);
> >         MqttDeliveryToken token = topic.publish(message);
> >
> >         // Wait until the message has been delivered to the server
> >         token.waitForCompletion();
> >
> >         // Disconnect the client
> >         client.disconnect();
> >         log("Disconnected");
> >     }
> >
> > Would I wait forever? Unacknowledged messages will stay forever in
> > internal in-memory queues and persistence?
> >
> > 2) Alternatively I might decide to not wait for completion and track the
> > confirms asynchronously in the callback:
> >         public void deliveryComplete(MqttDeliveryToken token) {
> >                 // Called when a message has completed delivery to the
> >                 // server. The token passed in here is the same one
> >                 // that was returned in the original call to publish.
> >                 // This allows applications to perform asychronous
> >                 // delivery without blocking until delivery completes.
> >
> >                 // This sample demonstrates synchronous delivery, by
> >                 // using the token.waitForCompletion() call in the main
> > thread.
> >         }
> >
> > I might want to track the in-flight messages in the application and
> > retransmit them if they are not acknowledged by a suitable timeout.
> > The spec allows for this by republishing the message using the original
> > message ID and setting the DUP flag in the MQTT PUBLISH header.
> > However this is not possible in Paho with the public API used in
> > Sample.java.
> >
> > We currently use our own MQTT client implementation but we would like to
> > switch to Paho because it will be actively maintained.
> > However our client autonomously performs retries.
> > Do you think it would be worthwhile to add this feature to Paho? What
> > would be the required effort?
> >
> > Thanks,
> > Ciao,
> >  Cristiano
> > ________________________________________
> > Da: paho-dev-bounces@xxxxxxxxxxx [paho-dev-bounces@xxxxxxxxxxx] per
> > conto di Nicholas O'Leary [nick.oleary@xxxxxxxxx]
> > Inviato: martedì 7 agosto 2012 10.03
> > A: General development discussions for paho project
> > Oggetto: Re: [paho-dev] Paho Java Client retransmission implementation
> >
> > Hi Cristiano,
> >
> > no, the paho client does not retry unacknowledged messages (except on
> > reconnect).
> >
> > Clients are not required to retry delivery of messages.
> >
> > Regards,
> > Nick
> >
> > On 7 August 2012 08:12, De Alti, Cristiano
> > <Cristiano.DeAlti@xxxxxxxxxxxx> wrote:
> >> Hi,
> >> Does Paho Java Client implement the retransmission of unacknowledged
> >> messages sent with QoS level 1 or 2?
> >> I'm browsing the code but I cannot find any references to a retry
> >> mechanism.
> >>
> >> Ciao,
> >>  Cristiano
> >>
> >> _______________________________________________
> >> paho-dev mailing list
> >> paho-dev@xxxxxxxxxxx
> >> http://dev.eclipse.org/mailman/listinfo/paho-dev
> >>
> > _______________________________________________
> > paho-dev mailing list
> > paho-dev@xxxxxxxxxxx
> > http://dev.eclipse.org/mailman/listinfo/paho-dev
> >
> >
> > _______________________________________________
> > paho-dev mailing list
> > paho-dev@xxxxxxxxxxx
> > http://dev.eclipse.org/mailman/listinfo/paho-dev
> _______________________________________________
> paho-dev mailing list
> paho-dev@xxxxxxxxxxx
> http://dev.eclipse.org/mailman/listinfo/paho-dev
>
>
>
> Unless stated otherwise above:
> IBM United Kingdom Limited - Registered in England and Wales with number
> 741598.
> Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU
>
> _______________________________________________
> paho-dev mailing list
> paho-dev@xxxxxxxxxxx
> http://dev.eclipse.org/mailman/listinfo/paho-dev
>
_______________________________________________
paho-dev mailing list
paho-dev@xxxxxxxxxxx
http://dev.eclipse.org/mailman/listinfo/paho-dev