Gang,
I have been toying with trying to run Polygene on a Flink cluster (no
success yet), and part of that is coming back to having a pluggable
Messaging/Event SPI, which is pluggable in the similar fashion to
persistence.
I am also getting a clearer point what is needed, and why we need a SPI and
not just one or more libraries for different messaging systems.
My initial (and possibly naive) thoughts are as follows;
It all comes down to UnitOfWork and the fact that Queue readers are in
their own threads.
1. A Consumer impl starts a UoW, starts a transaction on the queue*
2. Polls the Queue (potentially blocking)
3. Data read is deserialized**
4. ValueComposite is forwarded to all listeners***
5. Any kind of regular Polygene happens
6. Persistence might be invoked and part of the UoW
7. Producer impl might be called to forward a message to another queue/topic
8. Eventually returns to Consumer and the UoW.complete() happens, which
needs to handle the transaction commit to all individual parts, of
UoW.cancel() to rollback all parts.
* Does it need to start transaction on any destination queues as well?
** Several issue here;
a. Visibility from where?
b. Multiple types on same queue, how to resolve?
c. Data formats not compatible with ValueComposites, such as raw binary.
*** addListener or to all Services with some tagging?
Non-trivial I would think, but not excessive by Polygene standards. :-)
WDYAT?
Cheers
--
Niclas Hedhman, Software Developer
http://polygene.apache.org - New Energy for Java
