Matthew Swift a écrit :
Hi Emmanuel,
Hi Matthew,
comments inline
2) In Java, it's pretty natural to create objects through constructor.
This may be true for constructors which compose the constructor's
parameters (e.g. list of RDNs or RDN+DN) into a new object (DN), but
it is definitely not true for constructors which perform parsing or
type conversions. There are very many examples in J2SE:
* String.valueOf(...) -converting objects to Strings
* Integer.valueOf(int) - converting an int to an Integer
* Integer.valueOf(String) - parsing a String as an Integer
* plus other primitive Objects (Boolean, Byte, Char, Float, ...)
In fact, the use of static factory methods is a very common design
idiom and is strongly recommended in many texts including Joshua
Bloch's "Effective Java" (item #1). By using static factories with
well known names (e.g. valueOf) we are able to avoid creating
duplicate objects (item 4) if this proves useful (i.e. through the use
of a cache) and also enforce the singleton property for the root DN
(item #2).
In any case, I thinkw e should have the constructors. Even for Integer,
you can do new Integer(n).
3) I'm not sure that handling a cache for DN is a valuable trick, as it
leads to some contention and the need to manage this cache in a
muli-threaded environement (this has to be carefully evaluated)
I have done already and it results in a 30-40% parsing performance
improvement as well as a significant reduction in garbage. Having said
that, I'm not totally convinced that this gain is really worth the
extra complexity and old generation memory usage - especially
considering large server based applications may have 1000s of threads
- e.g using LDAP from within a servlet - all those thread local caches
could use a significant amount of memory!
Our implementation uses a small thread/schema local cache mapping
String values to DNs. In most applications I think that it's
reasonable to assume that 99% (perhaps more) DNs share the same parent
or grand parent. When parsing a DN we first check if the string is in
the cache and, if not, parse the RDN and recursively repeat for the
parent DN (in the SDK a DN is implemented recursively as an RDN+DN).
Well, I don't really think that it's anything but implementation
dependent, so from the API POV, it's irrelevant. As soon as we add the
valueof() methods, those who want to add a cache can do it.
I don't know why I raised this point...
The base constructor we can have are probably something like:
DN()
DN(String dnStr)
DN( RDN... rdns)
DN( RDN rdn, DN parent)
I like the DN( RDN rdn, DN parent) constructor - we support this via
an instance method called DN.child(RDN). I think I prefer the
constructor approach since it is not clear from our "child" method
whether or not it modifies this DN or creates a new DN. A constructor
is more obvious. You may want to have a concatenation constructor
DN(DN child, DN parent) for constructing DNs of entries within a
subtree using a relative DN (or "local name").
Why not a DN(DN child, DN parent) constructor two. It does not hurt and
can help.
One thing that is a bit tricky is whether or not the API should order
RDN parameters in little-endian (LDAP World) order or big-endian
(everyone else outside of LDAP) order. I think first time users may be
surprised by LDAP's unnatural little endian ordering.
I think we should keep the LDAP order when using DN( RDN...)
constructor. For instance, if we want to create "dc=example, dc=org",
that would be :
DN( "dc=example", "dc=org") (here, I use the String, but you should read
RDN)
Also, I strongly believe that DNs and RDNs and AVAs should be
immutable objects (as well as any other low level API type). What do
you think?
DN and RDN should be immutable, sure. AVA, I have some doubt.
Also, on the subject of AVAs - we have the AVA type as an inner class
in RDN. I'm not particularly happy with this this, but less happy with
it being a standalone class since AVAs are only used in RDNs and may
introduce confusion elsewhere. For example, filters also use attribute
value assertions but these are not the same type of object as an AVA
even if they have the same name. For example, AVAs (in RDNs) do not
allow attribute options or matching rules to be specified.
I don't really like inner classes in this case for two reasons :
- It will be a big class, and the RN class while be hundreds of line
long. Not cool
- If we just use an Inner class just because we want to hide it from the
other classes, then I think it's probably better to have it package
protected (ie, no qualifier for this class).
This can be discussed further, as your pont about its use in other
context can be figured out.
Thanks !
