[jira] [Commented] (BOOKKEEPER-312) Implementation of JMS provider

["Matthieu Morel (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/BOOKKEEPER-312?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13471644#comment-13471644 ] 

Matthieu Morel commented on BOOKKEEPER-312:
-------------------------------------------

I just added a few comments on the review board. ( https://reviews.apache.org/r/6277/ )

My view is that the effort is quite comprehensive, and missing features clearly identified. The patch should take the latest comments into account and be updated to match the latest trunk version. Then hopefully integrated.

Considering the current limitations are clearly outlined, it will be a good basis to build upon. Then we'd create specific tickets for each of the missing features, so that they are more focused.
                
> Implementation of JMS provider
> ------------------------------
>
>                 Key: BOOKKEEPER-312
>                 URL: https://issues.apache.org/jira/browse/BOOKKEEPER-312
>             Project: Bookkeeper
>          Issue Type: Sub-task
>            Reporter: Mridul Muralidharan
>             Fix For: 4.2.0
>
>         Attachments: hedwig-client-jms.patch, hedwig-client-jms.patch.1, hedwig-client-jms.patch.2, hedwig-client-jms.patch.3, hedwig-client-jms.patch.4, hedwig-client-jms.patch.5
>
>
> The JMS provider implementation conforming to the 1.1 spec.
> The limitations as of now are :
> 1) No support for Queue's : Hedwig currently does not have a notion of JMS queue's for us to leverage.
> 2) No support for noLocal : Hedwig DOES NOT conform to JMS model of connection -(n)-> session -(n)-> publisher/subscriber. Each session has a hedwig connection.
> Currently I am simulating noLocal, but this IS fragile and works for the duration of connection - ONLY until the message id is still in a LRUCache. As mentioned before, this is a kludge, and not a good solution.
> 3) Note that everything is durable in hedwig - so we do not support NON_PERSISTENT delivery mode.
> 4) Calling unsubscribe on a durable subscription will fail if it was NOT created in the current session.
> In hedwig, to unsubscribe, we need the subscription id and the topic ... 
> To simulate unsubscribe(), we store the subscriberId to topicName mapping when a create* api is invoked. Hence, if create* was NOT called, then we have no way to infer which topic the subscription-id refers to from hedwig, and so cant unsubscribe.
> The workaround is - simply create a durable subsriber just as a workaround of this limitation - the topicName will be known to the user/client anyway.
> 5) Explicit session recovery is not supported.
> Reconnection of hedwig session (either explicitly or implicitly by underlying client implementation) will automatically trigger redelivery of un-acknowledged messages.
> 6) Because of the above, setting the JMSRedelivered flag is almost impossible in a consistent way.
> Currently, we simulate it for redelivery due to provider side events : rollback of txn, exception in message listener (primarily).
> At best we can simulate it with a kludge - at risk of potentially running out of resources ... this is being investigated : but unlikely to have a clean fix.
> 7) Hedwig only supports marking all messages until seq-id as received : while JMS indicates ability to acknowledge individual messages.
> This distinction is currently unsupported.
> 8) JMS spec requires
>     "A connection's delivery of incoming messages can be temporarily stopped
> using its stop() method. It can be restarted using its start() method. When the connection is stopped, delivery to all the connection’s MessageConsumers is inhibited: synchronous receives block, and messages are not delivered to MessageListeners."
>   We honor this for undelivered messages from server - but if stop is called while there are pending messages yet to be delivered to a listener (or buffered in subscriber for receive), then they will be delivered irrespective of stop().

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira