Please review draft JEP: JMX Specific Annotations for Registration of Managed Resources

Wed Mar 4 17:42:56 UTC 2015

Thank you for updating the JEP text referencing JSR 255.

Perhaps unsurprisingly I disagree with many of the differences between
this proposal and the one we carefully thought out in JSR 255. Even
though a lot has changed in the meanwhile, I don't see anything that
invalidates our assumptions of the time.

For reference, a snapshot of the JSR 255 javadoc is at
http://hg.openjdk.java.net/jmx2/jmx2/file/f417598a7b72/javadoc.
Unfortunately, it appears that the Mercurial server now serves these
files as application/binary, probably because of the change here:
http://mercurial.selenic.com/wiki/UpgradeNotes#A1.9.1:_guessmime.2C_revert_behavior_restored.

To address some specific points:
> would you care to elaborate what you find to be not "even correct Java"?

As of Java 8, annotation elements cannot have null values so the
"default null" clauses are nonsense. I have not seen any proposal to
change this in Java 9. The @ManagedBean definition also has an obvious
syntax error.

> - ability to annotate fields turning them into attributes

This might be useful for read-only attributes. I'd question whether it
is useful for read/write attributes, because I think it will be very
unusual for you to want neither validation of the new value nor
behaviour to be triggered by the set.

> - ability to provide metadata directly in the annotations, not relying
> solely on inferring them from the annotated element

I'm not sure what specifically this refers to. Do you mean for example
that it is possible to add @ManagedAttribute to a method that does not
look like getFoo() and nevertheless have the annotation say that the
attribute is called foo? I don't see any particular advantage to that
flexibility. The getFoo() pattern is already familiar, and having a
second, completely different way of specifying the name just
complicates the spec for not much benefit.

> - missing @ManagedConstructor to expose a constructor

We deliberately omitted this. The fact that MBeanConstructorInfo
exists at all is in my opinion a mistake in the original JMX
specification. What does it mean for an MBean to tell you how to
construct another instance of itself? And if the purpose is to specify
which constructors from this class are available to the MBean Server,
there's no use for names and descriptions, and there's no particular
advantage over just saying that all public constructors are available.

> - optional 'service' argument to @ManagedBean annotation which will be
> reflected in the descriptor's 'interfaceClassName' field to provide a
> guidance about the recommended service interface when using
> JMX.newMXBeanProxy()

If you have such an interface, why wouldn't you just use it to define
the MBean and dispense with annotations?

Some other comments:

* @ManagedBean.

We called this @MBean because we also had an @MXBean annotation. That
annotation exists today, but JSR 255 allowed it to be applied to a
class as well as to an interface. It appears that @ManagedBean only
defines MXBeans, which is a legitimate choice but, first, it should be
called out more explicitly, and, second, wouldn't it then make sense
to extend the existing @MXBean annotation?

The specification is inconsistent as to whether the annotation is
@ManagedBean or @MBean.

I think it is a fair idea to have an objectName field, but the idea of
randomly appending numbers to the name for disambiguation is broken.
Something like @ObjectNameTemplate from JSR 255 is more appropriate.

The text for the notifications() member references @TypeMapping but
does not say what that is. The declared type is Notification[] and the
text defines an annotation @Notification, but there is already a class
called Notification in javax.management.

I think that the simple "name=value" syntax used by JSR 255's
@DescriptorFields is preferable to the unspecified and verbose type
@Tag. I don't see an advantage to making people write @Tag(name =
"foo", value = "bar") rather than just "foo=bar". This syntax is
already present in the JMX spec, for example in the
ImmutableDescriptor constructor.

* @Notification.

As I mentioned, you can't use that name.

The first paragraph of the description is indecipherable.

The NotificationSender interface is unspecified. Based on the example,
I think it could potentially be a major usability improvement but it's
hard to be sure.

I think it's extremely ugly to propagate the misspelling clazz into an
API where people will have to write it. Also, if it must extend
Notification then it should be specified as Class<? extends
Notification>.

* @ManagedAttribute

It's extremely strange for this to have getter and setter fields. Why
wouldn't it just be applied to those methods?

Promoting units from a descriptor field to a separate annotation
member seems like a good idea. The specified value would be copied
into the Descriptor.

* @ManagedOperation

I don't see any reason to allow the name to be different from the
method name. It just complicates the spec.

Instead of repeating a description member inside every annotation, JSR
255 defined a top-level @Description, which included elements for
internationalization. Hardcoded strings are a step backwards.

Defining Impact inside this annotation is questionable. I'd expect
user code and possible future API changes to want to reference it
independently of the annotation. Also, the JSR 255 enum Impact had
methods to convert to and from the integer codes used by
MBeanOperationInfo.

* @ManagedParameter

The text repeatedly says operation name and method name when it means
parameter name. I assume that if the name member is empty then the
parameter name from reflection is used, which since Java 8 could be
the actual name of the parameter if the class was compiled with
-parameters.

* @RegistrationHandler

It seems like an API smell for an annotation to say that it must be
applied to methods with a certain signature. I think a much better
approach would be to change the existing MBeanRegistration interface
so that its methods have default implementations that do nothing. That
removes the main reason that this interface is a pain: having to
implement four methods when you're usually only interested in one. You
could also add a preDeregister overload with MBeanServer and
ObjectName parameters, again with a default implementation.

Éamonn

2015-03-04 0:47 GMT-08:00 Jaroslav Bachorik <jaroslav.bachorik at oracle.com>:
> On 4.3.2015 02:09, Eamonn McManus wrote:
>>
>> Could you explain what you mean by this, regarding the annotations
>> that were already agreed on by the JSR 255 Expert Group:
>>
>> * Smaller scope compared to the proposed solution
>
>
> This is a leftover from the previous proposal which had wider scope than
> what is presented now. Still a few points remain.
>
> - ability to annotate fields turning them into attributes
> - ability to provide metadata directly in the annotations, not relying
> solely on inferring them from the annotated element
> - missing @ManagedConstructor to expose a constructor
> - optional 'service' argument to @ManagedBean annotation which will be
> reflected in the descriptor's 'interfaceClassName' field to provide a
> guidance about the recommended service interface when using
> JMX.newMXBeanProxy()
>
>> * Conceptually in pre JDK7 era
>
>
> I am afraid this relates more to the implementation - or at least the code I
> was able to reconstruct from the repo history. Shouldn't be mentioned here.
>
>>
>> I have a number of other comments, but procedurally I'm not sure what
>> the precedent is for summarily discarding work previously done in the
>> JCP on the same subject. I'd certainly have expected this JEP to start
>> from that work, rather than proposing a starting point that isn't even
>> correct Java.
>
>
> Well, this is a draft review. The JSR 255 annotations work is not discarded.
> It is mentioned in the alternatives. The purpose of the open review is to
> find out whether it is ok to continue with proposed feature, modify it to
> use eg. JSR 255 work or abandon it completely.
>
> -JB-
>
>
>>
>> Éamonn McManus, former JSR 255 Spec Lead
>>
>> 2015-03-03 8:27 GMT-08:00 Jaroslav Bachorik
>> <jaroslav.bachorik at oracle.com>:
>>>
>>>
>>> Hi all,
>>>
>>> Please review this draft JEP for JMX Specific Annotations for
>>> Registration of Managed Resources:
>>>
>>>       https://bugs.openjdk.java.net/browse/JDK-8044507
>>>
>>> Background:
>>> Current mechanism of defining an MBean requires to provide an MBean
>>> interface and its implementation. The interface and the implementation must
>>> conform to the strict naming and visibility rules in order for the
>>> introspection to be able to bind them.
>>>
>>> At least the same level of verbosity is required when adding an MBeanInfo
>>> to generate MBean metadata.
>>>
>>> All this leads to a rather verbose code containing a lot of repeating
>>> boilerplate parts even for the most simple MBean registrations.
>>>
>>> This JEP proposes to add a set of annotations for registration and
>>> configuration of manageable resources (in other word 'MBeans'). These
>>> annotations will be used to generate all the metadata necessary for a
>>> resources to be accepted by the current JMX system.
>>>
>>> Thanks,
>>>
>>> -JB-
>
>