Hi Emmanuel,
I totally agree with your thoughts regarding the Control API. As per
usual, I prefer getOID to getOid... :-)
Our OpenDS SDK Control package is not yet finished and needs some
serious clean up. In fact, I was planning to turn our Control abstract
class into an interface.
For cases where client apps want to use a control for which there is no
existing sub-class implementation we provided them the option of using
the "GenericControl" class instead of being forced to implement a
sub-class. The GenericControl class is pretty straightforward - it
implements Control and provides three constructors:
GenericControl(String oid) // non-critical, null value
GenericControl(String oid, boolean isCritical) // null value
GenericControl(String oid, boolean isCritical, ByteString bytes)
I find the GenericControl name a bit non-obvious. An alternative
approach is to hide the class (package private) and expose the
constructors using a separate "Controls" utility class. E.g:
public static final class Controls {
public static final Control newControl(String oid) { return new
GenericControl(oid); }
}
I think that we'll probably do this in the OpenDS SDK since we'll also
need at least one other utility method so that we can provide immutable
request/response wrappers:
public static final Control unmodifiableControl(Control control) { ... }
Note that the Control interface is immutable, but that does not stop
implementations from being mutable.
The Control API must distinguish between null values (i.e. value not
present) and empty values (i.e. value present but zero-length).
I don't know if it is out of scope for now, but do we want to support
extensibility? In particular, how can client applications implement
their own controls? There are two main issues here that I have encountered:
1. Decoding of response controls: if I have a response control whose
type is "MyControl" do I want the LDAP SDK to automatically decode
it to this type? Or will the client application have to do it.
Here's some pseudo code to illustrate my point:
// Result contains a MyControl response control.
Result result = connection.add(entry);
// Option #1: Uses an internal map of Control implementation
classes -> OID + decoder
MyControl control = result.getControl(MyControl.class);
// Option #2: Uses an internal map of OID -> decoder
MyControl control = (MyControl) result.getControl(MyControl.OID);
// Option #3: No internal map - client has to manually decode
Control control = result.getControl(MyControl.OID);
MyControl myControl = new MyControl(control);
I prefer the first approach for simplicity but it requires a
public API for registering Control implementations (as does option
#2) or use introspection and require that all implementations
provide an OID field and a constructor having a single Control
argument. Option #3 is quite verbose for clients IMO.
I think that it's safer if the request/response API decodes the
Control each time rather than caching the decoded control. This
will make it possible to support immutable request/responses.
If it sounds like I getting ahead here, the point of this issue is
that if we want to provide an simple decoding mechanism like #1
then we will need to have some way for the SDK to be able to
decode the Control. This means either having a registration
process, or using introspection and having a well defined
constructor and OID field.
The same problem will present itself for the following API features:
* decoding extended responses
* decoding intermediate responses
* decoding request controls (server-side so out of scope)
* decoding extended requests (server-side so out of scope)
2. Encoding/decoding controls: many control values are encoded using
ASN1. Do we want to provided ASN1 support? This will also apply
for new extended operations.
I think that these questions are only applicable if we decide to support
extensibility.
Matt
On 25/01/10 01:30, Emmanuel Lecharny wrote:
Hi guys,
as I'm working on the messages, I have looked at the Control
Interface. Here is the way it's used in all the different APIs :
ADS :
we used the JNDI interface, and we will change to switch to a LDAP API
Control
jLdap/LDAPSdk
=============
class :
LDAPControl
constructors :
LDAPControl()
LDAPControl(String, boolean byte[]))
methods :
getID()
getValue()
isCritical()
newInstance(byte[])
clone()
JNDI
====
interface :
Control
methods :
getEncodedValue()
getID()
isCritical()
ODS
===
abstract class :
Control
constructors :
Control( String, boolean)
methods :
getOID()
getValue()
hasValue()
isCritical()
UID
===
class :
Control
constructors :
Control(String)
Control(String, boolean)
Control(String, boolean, ASN1OctetString)
methods :
encode()
equals( Object )
getControlName()
getOID()
getValue()
hasValue()
isCritical()
Checking all those implementations, I would suggest that we stick with
something like :
interface Control
<discussion>
Not sure we need a class, when most of the controls will be named,
like VLVControl, PageSearchControl, ...
ODS define an abstract class, which means nobody can create an
instance of it. At this point, I think it's probably better to go for
an interface, hidding the asbtract class to the client.
</discussion>
constructors :
<discussion>
No need of them, if it's an interface. From the server POV, this will
be an issue, as we must be able to create Controls while decoding them
and we don't necessarily know if the control will be supported or not.
We will probably need an intermediate ControlImpl and an
AbstractControl class in order to deal with this, but from the Client
POV, this is nt relevant
</discussion>
methods :
getOid() or getOID().
getValue()
isCritical()
<discussion>
Those three methoda re the bare minimum we need from the client side.
I'm not sure that the hasValue() method is necessary, when the
getValue() returning a null value will provide the same result.
</discussion>
Thoughts ?
