On Feb 7, 2007, at 10:51 AM, David Jencks wrote:
On Feb 7, 2007, at 9:54 AM, Dain Sundstrom wrote:
On Feb 7, 2007, at 8:01 AM, David Jencks wrote:
The problem is our users' applications will make use of mapped
name in other application servers, and since every application
server that I know if is implementing these with Global JNDI, it
becomes a defacto standard and requirement for Jee5. Moreover,
I believe that our GlobalJNDI names must be simple normal names
(i.e, not encoded abstract names) you would see in other
application servers, because users will annotate their code with
the mapped names, effectively locking in the Global JNDI names
that they expect to work in our application server.
umm, that assumes that either every other app server has come up
with exactly the same scheme for global jndi names so they are in
fact interoperable or that we can imitate everyone elses naming
scheme at once.
um, I'm not making that assumption. I'm assuming that other
systems are using Global JNDI for resolving refs and that users
will have "normal" JNDI names hardcoded into their apps. By
normal names I would say that they are simple words separated by a
'/' character.
I'm saying:
1. simple names are good
2. simple names have nothing necessarily to do with global jndi
althought that is one possible implementation strategy
3. if users expect portability across app servers we need to do
some actual research on what other app servers do and find out if
there is in fact a de-facto standard. Hand waving won't cut it.
I did do resarch, but the data will not be satisfactory to you. The
only server with formal docs on this is Glasfish and they use Global
JNDI as I described. As for the other servers, they have posts to
mailing lists that clearly imply that they have implemented mapped
name the same way that Glassfish did. Alternatively, I can not think
of a single appserver other then Geronimo that doesn't make extensive
use of Global JNDI.
So do you agree that Global JNDI is the defacto required
implementation for these and other similar entries?
no.
Since the beginning of geronimo we've carefully stayed away from
relying on global jndi for resolving references since it imposes
global constraints on what you can deploy at once in the app
server, despite every other app server I know of relying on
global jndi for resolving references. I'm extremely reluctant to
abandon the lack of conflicts between apps that we have now to
run after an alleged similarity with other app servers without
thorough investigation of compatibility between other app servers
and thinking about other choices that would preserve the lack of
conflicts.
I understand your reluctance and know the history, but reality is
*everyone* uses Global JNDI. I challenge you to find a single JEE
server that doesn't. Since everyone does, it is the defacto
standard, and it is becoming more ingrained into the specs. I
think we need give up on our custom system, and simply implement
it the way everyone else dose.
I don't really understand what your insistence on global jndi has
to do with the problem of resolving these annotations. The user's
app isn't directly using jndi for this. To me you are saying "we
can't do anything better than anyone else because it might be
different".
But they do use Global JNDI also. We have tons of posts to the users
list from confused new users that can't figure out how to specify
Global JNDI names for data sources. I believe that the introduction
of mappedName in JEE5 is the proverbial straw, and we should just use
Global JNDI and be done with it.
Note that the use of any particular style of name in such
annotations does not imply that the target is actually bound in
jndi: all it requires is that we can find the resource somehow.
Two alternatives that I would prefer to global jndi are:
1. We know the type of the thing we're looking for, so we can
simply treat the provided string as an (extended) ejb-link,
resource-link, etc and search the ancestor tree of the current
app for a unique match. IMO this would be a lot simpler to
implement that relying on global jndi, because we already have
the code implemented and don't have to bind anything anywhere.
2. "scoped" global jndi. Each application gets a "global" jndi
tree that only includes stuff from itself and its ancestor
graph. This avoids conflicts and should satisfy those with a
jndi fetish.
There is nothing wrong with either proposal, and either would be a
step forward. I just think we shouldn't invent something new and
just give users what they expect.
Do you think (2) would surprise people too much? I'm very
reluctant to abandon the conflict avoidance of what we have in
order to copy the lowest common denominator. (1) is the easiest
solution I can think of since it's basically already implemented
and doesn't involve coming up with a binding strategy for
everything and (2) seems to me to combine the advantages of global
jndi with the conflict avoidance we have now. Why is plain global
jndi better?
I am not saying that Global JNDI is better. I'm saying it is the
norm. People are used to the problem of JNDI and have procedures to
avoid naming conflicts. When they have naming problems
(NameNotFoundException) they have tools (e.g., JNDI browser, JNDI
dump, etc) to debug the problem and known how to fix it (change the
name).
-dain
