Can acknowledgements be asynchronous for auto_ack?

[timfox]

It's my understanding that NACKs (cancels) can always be sent asynchronously from the client to server, since it won't affect the reliability guarantee if they are lost for some reason.

However, with ACKs then, if the ACK fails, the client needs to know this, in the case of CLIENT_ACKNOWLEDGE, or transactional (where the commit does an implicit ack of the messages), therefore the call to acknowledge needs to be synchronous.

However with DUPS_OK, duplicates may happen anyway, therefore we can get away with asynchronous acks, since if one is lost this may result in redelivery which we are ok with anyway.

What about AUTO_ACK? Does the ack need to be synchronous in this case?

With auto_ack without a message listener, the message is acked before it is processed. So messages could be lost anyway, if the client fails. So the once and only once does not apply anywhere. Since the reliability guarantee is lower is it acceptable to use asynchronous acks with AUTO_ACK as well?

[adrian.brock]

No. Consider the following:

Synchronous:
Client: receive(); -> server
Client: AUTO_ACK -> synch
Client: process message (only once)
Client: Crash
Client: Reboot
Client: Cannot re-receive the message

Asynchronous:
Client: receive(); -> server
Client: AUTO_ACK -> asynch
Client: process message
Client: Crash (the asynch ACK never made it to the server)
Client: Reboot
Client: Can re-receive (and reprocess!) the unacked message

In fact, part of the contract of receive() for AUTO_ACK is that
the ACK has been successfully done such that when the client
does processing on the message, it knows that it won't get the message again.

[timfox]

In fact, part of the contract of receive() for AUTO_ACK is that
the ACK has been successfully done such that when the client
does processing on the message, it knows that it won't get the message again.

Thanks for the clarification.

Right, this is a "at most once" guarantee I guess, since the message might be lost if failure happened between acking and the client actually getting the message.

[timfox]

BTW, the reason I wanted to check this, is that a certain very popular open source messaging system (you know the one I mean ;) ), in it's out of the box configuration, batches up acks in AUTO_ACK_MODE - so basically AUTO_ACK has the same delivery guarantee as DUPS_OK, which seems somewhat disingenuous, but probably helps them on benchmarks.

Which leads to the question of whether our out of the box configuration should be JMS spec compliant, or optimised for performance and the user has to actively set a flag --JMScompliant. Currently we are out of the box JMS spec compliant.