Interview: Gregor Hohpe on Conversation Patterns and Cloud Computing
In a new interview, recorded at QCon London, Google architect Gregor Hohpe talks to Stefan Tilkov about integration and conversation patterns. Gregor, who co-authored the classic ”Enterprise Integration Patterns”, is also well-known for his ”Starbucks Does Not Use Two-Phase Commit” article, in which he showed the difference between the kind of co-ordination computer scientists wish for as opposed to what is found in practice.
Gregor talks about the relevance of his first patterns work for today’s SOA scenarios, and offers some first thoughts on his new work on “conversation patterns”. Building upon the earlier, more basic patterns, he sees them as playing a critical role in real-world interactions, with analogies in the natural world. Another topic is description languages for conversations, specifically WS-BPEL and WS-CDL.
Gregor also talks about the similarities and differences in several approaches to cloud computing. He contrasts low-level services to high-level offerings, and notices similar patterns in different companies' offerings based on experience gained in the "Cloud Computing" track at QCon he chaired.
Watch the full interview (41 minutes).
Agree in part, definitely leaves room for improvement IMHO
I definitely agree that "it should not be one big transaction". However, I would consider as transactional at least those parts of the system tied together synchronously. For instance: consuming messages from a queue and processing the results in a database. Or: checking a balance and crediting an account.
The proposed model to solve transactional problems in a business way (trusting the goodwill of your bank) clearly does not work: how will the bank find out that they lost my 50 bucks? Either they find out themselves (pre-supposing some transaction monitor to be in place) or they know because I complain (in which case the damage is already done).
A similar thing happened to me recently with Amazon (see my blog post). I consider myself a good customer of theirs, yet I had no refund so far :-(
If you follow my rule of thumb then the scope of a transaction will always be very limited:
-2 systems in queued order processing
-N systems in SOA/workflow processing, where N is the number of services synchronously invoked/connected
I think these are very acceptable. Again, I do not claim there should be one big transaction, but transactions should be used where they are of value.
Atomikos - 3rd generation transaction management for XTP and SOA
PS I also believe Starbucks would offer better service if they used two-phase commit to process their queued orders :-)