QCon London 2008: Mark Little "Diary of a Fence-sitting SOA Geek"
These are my unedited notes from Mark Little's talk at QCon London 2008.
- Observes that the talks have been sort of an anti-climax
- Some history about distributed computing (skipping some here)
- In the late 1960s/1970s, spending months to integrate four machines
- RPC invented in the early 1970s
- RPC is NOT dead, neither is it pining for the fjords
- RPC still running a lot of backend systems
- 1990, some physicist invented the World Wide Web
- Back then: "What does he know that we don't? How dare he? How could he come up with something when he's not even a computer scientist?"
- 1999, Web Services were invented
- CORBA's biggest problem: Microsoft not involved
- Firewall vendors won't open up to CORBA or DCOM (dynamic ports), so we'll show them
- HTTP is a useful 'transport' after all
- 2000-2005: spent 5 years going backwards to distribution transparency - it feels like 1970 all over again
- Distribution must be explicit, otherwise you're not able to deal with failures
- Same fundamental laws, for Web services, WWW, CORBA, DCOM:
- Develop "entity", define the unit of work it supports
- Search for the entity
- Request entity to perform unit of work
- maybe try to make the remote interaction local
- maybe do some "enterprise" stuff as well (security, transactions, replication)
- Different approaches
- EDA
- SOA
- Message-oriented (MOM, Pub/Sub)
- Tuple Spaces
- RPC
- sync/async
- group communication
- There is no such thing as a global panacea!
- Problems with tight coupling introduced by trying to make a remote interaction look like a local interaction
- Explanation of client/server stubs/skeletons
- Definition of SOA: Not purely technology, an architectural style trying to achieve loose coupling
- Services as re-usable units
- Explicitly not CORBA w/ angle brackets, coarse-grained services
- SOA has two architectural constraints:
- A small set of ubiquitous interfaces to all participating agents, only generic semantics -- often based on specific, but doesn't have to be
- Descriptive messages, described by schema
- sounds vaguely REST-like
- At one extreme: very specific service interfaces, akin to IDL
- At the other extreme: (e.g. single operation doWork)
- There are degrees of coupling (not intended to evoke sexual associations ;-) and you should choose the level that's right for you
- Shows picture of OSI 7 layer model (only known to people with grey hair)
- With a uniform interface, your application has to do more
- The same requirements are present throughout the stack
- split differently between the application and the infrastructure
- uniform interface enables generic infrastructural asupport
- specific interface allows for more limited generic support
- developers have to do less
- REST over HTTP
- original HTTP 0.9 had only GET
- now HTTP has 8 methods
- HTTP is NOT a transport
- Often stated that HTTP/REST is only suitable for hypertext
- Weak argument: nothing inherently limiting machine usage
- Extra infrastructure could help (anyone remember URN name servers 1994/1995?)
- Google and other search engines are fantastic name servers
- Not enough "application" standards -- good point
- Lack of good tooling (changed, e.g. via JAX-RS, WCF)
- What about Web services?
- Began life as CORBA-over-HTTP
- Just as easy to develop non-SOA apps in WS as it is to write SOA apps with other tooling
- WS-Addressing and WS-RF both good examples of WS not being used for SOA
- Web Services and SOA gives benefits
- Enterprise realities
- Customers want interop, guaranteed delivery of messages, transactions (not ACID ones), audit trails, bullet-proof security, machine-readable contracts
- Shows innoQ's WS poster ;-)
- Can convey the wrong image -- that you need everything
- Standards: The Web is a series of standards, universal adoption has to count for something, REST/HTTP is ubiquitous
- application semantic interop is missing
- Shows REST transactions by HP in 2000 (!)
- Back then, Mark Baker's continuous interruptions in the mailing lists led to this
- GET on http://\
/transaction-coordinator - PUT on http://\
/transaction-coordinator/begin?clientid will start new tx and return http://\ /transaction-coordinator/\ - Misses standards e.g. for search (ed: what about OpenSearch?)
- WS-* has driven portability
- clearly defined semantics for transactions, security, reliable messaging -- not specific to HTTP
- WS-* vs. REST
- Internet-scale computing for WS is no longer being talked about
- What can REST learn from WS-*: Uniform interface isn't enough, e.g. support for transactions, in a standardized manner
- Popularize HTTP outside the browser
- Just because the Web works doesn't make it right for everything (when did you last ride a horse to work?)
- Don't abuse transports, they don't like it -- HTTP is NOT a transport
- Adopt SOA principles within Web services (no more WS-RF, WS-Context instead of WS-Addressing)
- Late binding is good
- KISS and make up
- Remember the Web uses HTTP
- One size doesn't fit all
- Use the right tool for the right job
- Both the REST and the WS communities should look at the work that has been done in the 80s and 90s
- Question: What would you say to people who say "Squid is the ESB"?
- Answer: It might work if you're just looking for an EAI framework, but he thinks an ESB is more than that