[
https://issues.apache.org/jira/browse/FELIX-227?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12478112
]
Steven E. Harris commented on FELIX-227:
----------------------------------------
I've used Spring, and maybe I missed some of its features, but its dependency
injection worked by way of supplying constructor parameters to or calling
methods on my classes, so its interaction with my class always respected the
encapsulation boundary.
When you write, "you don't have to focus on which value a container-managed
field has", this requires one to know which fields are container-managed. Maybe
that's what I'm struggling with. I was trying to take an existing class and use
it with iPOJO's property-injected fields, and I obviously ran into a lot of
trouble. If I had started with the assumptions you're describing, I could have
written the class differently.
But consider what would happen if I decided to change which fields should be
bound to service properties. The class itself keeps on compiling just fine, but
now it may be broken, failing to initialize fields no longer managed by the
container. It's hard to be comfortable with this kind of code.
Your "method" attribute on <property> is almost exactly what I was asking for,
but I would expect that the "field" attribute would be mutually exclusive with
the "method" attribute. If I want iPOJO to call on my methods, I probably don't
want it actually touching my fields.
However, a few other questions arise:
Should an updated factory-based Configuration still cause an instance to be
killed and recreated, or should you assume that an instance can tolerate being
updated dynamically if it requests these "method" callbacks? Note that DS
deactivates and reactivates an factory-created instance when its Configuration
changes. I'm not sure which is the right way to go.
Also, what about constructors? Would these property callbacks be called after
the constructor runs? In my report above, I noted that initializing these bound
fields in the constructor mistakenly overwrites the iPOJO-property-bound
values. What I'd prefer is to allow the default constructor (or one taking a
BundleContext) to run, with its fields initialized by the constructor code,
unmolested by iPOJO, and then have the fresh instance be called on with each of
its "method" property callbacks.
> iPOJO should allow configuration and service properties to be bound via
> setter/getter methods, not just via direct fields
> -------------------------------------------------------------------------------------------------------------------------
>
> Key: FELIX-227
> URL: https://issues.apache.org/jira/browse/FELIX-227
> Project: Felix
> Issue Type: Improvement
> Components: iPOJO
> Environment: Not relevant.
> Reporter: Steven E. Harris
>
> iPOJO's binding of configuration properties and service properties directly
> to component fields causes a few difficult situations that could be eased by
> optionally binding the reading and writing to getter and setter methods
> instead.
> First, consider a component class with invariants that span multiple fields.
> iPOJO can change the value of a field without the owning class knowing,
> depriving it the chance to update dependent fields that participate in the
> invariant. Examples include recalculating and caching an expensive result
> whenever some input field changes, or validating a changed value and updating
> other fields in response.
> Second, at present iPOJO sets the value of bound fields before the component
> class constructor runs. If the constructor attempts to initialize some
> fields, it may be inadvertantly overwriting the initial values supplied by
> iPOJO. Writing the constructor to deliberately ignore values that might be
> bound by iPOJO flies in the face of the very name of the project: a POJO
> isn't supposed to know that it's being silently manipulated like this, and
> hence a POJO should be written in standard form: Initialize variables to sane
> defaults, including constructor parameters, expecting them to be overwritten
> later.
> Take for example a class that has two fields, one of which is bound to a
> configuration property:
> public class Example {
> private String bound;
> private String dependent;
> public Example() {
> bound = null;
> dependent = "empty";
> }
> public void setBound(String s) {
> bound = s;
> dependent = null == s || 0 == s.length() ? "empty" : "full";
> }
> public String getDependent() {
> return dependent;
> }
> }
> If written in this manner, the constructor mistakenly overwrites the initial
> iPOJO-provided value for "bound" by initializing it to null. But to resist
> initializing "bound' is also dangerous; how would one reading this code have
> any idea that "bound" might get set to a different value before the
> constructor runs, or while the instance is live?
> Also, consider that if "bound' changes silently once the instance is live,
> "dependent" will fall out of step, as the invariant maintained in setBound()
> can be violated.
> If iPOJO would allow property binding to optionally work by way of getter and
> setter methods, one and a half of these problems could be avoided. The
> missing half relates to construction. If we ask iPOJO to defer setting the
> "initial values" until the constructor completes, we may have to defer some
> initialization that would use the values not yet available.
> Trying to write a POJO class that gets manipulated on the sly by iPOJO is
> proving to be more tricky than just writing some of the ManagedServiceFactory
> code myself, as I'm forced to adjust my would-be POJO service class to deal
> with these weird initialization and invariant maintenance problems normally
> solved by member variable encapsulation. Perhaps we should look at Spring's
> example that better acknowledges not just the technical possibilities, but
> the logical difficulties in iPOJO's kind of silent injection and manipulation.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
