> But if your
> > application _does_ obey REST principles (which is what is implied by
> > "doing REST"), then, yes, it surely does exhibit the properties stated
> > earlier.
>
> No... it MIGHT exhibit them. You can obey the principles of REST and
> still create a pig.

Sorry, that's wrong.  A system that abides by the constraints enumerated 
in the REST style will evince the properties associated with that 
constraint relative to a system that does not.  To use the caching 
example again:  There is a fixed cost in running the algorithm that 
produces a message and there is a fixed cost in transmitting that 
message over a network.  If the algorithm is called twice, those costs 
are assumed again.  If the message is cached, however, the algorithmic 
costs can be eliminated for future requests (until the cache gets stale) 
resulting in better performance.  Now, one could write a really bad 
algorithm that runs in time X, and another person can write the same 
algorithm that runs in time considerably less than X, but the algorithm 
is outside the scope of REST.  All other things being equal a system 
that employs caching is more performant than one that does not.

I have a feeling that you'll jump on the "all other things being equal" 
bit, but see here: http://en.wikipedia.org/wiki/Ceteris_paribus.  How 
else to analyze and predict the properties of complex systems?

> It will also be either completely useless (if there really is no
> state) or insecure and unreliable (if all state is retained on the
> clients). Statelessness of interaction is one thing, having no state
> at all isn't viable (if I place an order I expect you to remember it).
The statelessness constraint is _solely_ about session (interaction) state.
> A level of statelessness helps in lots of cases, but it doesn't mean
> the service is now reliable, just that there are less things to go
> wrong (which is a good thing).
Exactly!  Less things to go wrong means higher reliability. 

To be fair, and as recognized in the REST dissertation, each constraint 
comes with a cost as well.  For instance, statelessness, as you noted, 
has a performance penalty in that information that could conceivably be 
kept on the server now has to be in the message.  To handle these 
conflicts one has to (and Roy did) determine the level of impact that 
the cost will have on the properties of the system you're trying to 
bring about.  If the cost is negligible, then it can be ignored.  If 
not, then it would be nice to offset it as much as possible.  In the 
case of statelessness, not only does statelessness offer greater 
reliability but also visibility to external monitoring and scalability, 
because state no longer has to be kept or managed (or shared) on the 
server(s).  Eliminating statelessness then has serious negative 
implications.  On the other hand the overall performance impact can be 
mitigated by introducing caching,  Since most systems are heavily biased 
towards reads, the caching constraint more than makes up for what's lost 
due to the statelessness constraint, resulting in a high-performing 
system overall.

- Pete

Reply via email to