On Thu, Jan 6, 2011 at 4:49 AM, Emmanuel Lecharny <elecharny@gmail.com> wrote:
On 1/6/11 2:36 AM, Alex Karasulu wrote:
Hi all,

Excuse the cross post but this also has significance to the API list.

Problem
------------

For our benefit and the benefit of our users we need to be uber careful with
changes after a major GA release. We have another thread where it seems
people agree with the Eclipse scheme of versioning and this sounds really
flexible for our needs. We can do a 2.0.0-M1 release at any time without
clamping down on API's. Only when we do a RC do we have to freeze changes to
interfaces.

The debate still remains as to what constitues an interface. Emmanuel seems
to disagree with configuration, schema, and partition db formats as being
interfaces of concern but for the time being we can just discuss those we do
agree on. There's no doubt about APIs and SPIs.

I don't disagree with Schema, but Schema are clearly defined by RFCs, there is no possible interpretation about their syntax and definition.

Absolutely I do agree with you. I was thinking there's bug or mistake we find with our core published schema or we change the ApacheDS meta schema. In this case we'd need to bump up to a major version because then just a software update will not solve problems with already created entries on disk using the older schema. The situation will be undefined - very hard to predict.
 
However, the schema manipulation API is in the scope of this discussion.


This is part of the LDAP API and is as critical as Dn, or Entry since it's tied together.
 
Partition and configuration are not part of the Ldap API, thus are irrelevant in this discussion about shared refactoring.


Right this has nothing to do with shared APIs but is relavent to the server. The same policies in API maintenance in shared will have to be applied to the server.

To me anything exposed is something to consider for backwards compatibility, not talking just API here. Whether it's LDAP extended operations, web services, or database formats, these things impact backwards compatibility. 
 

Solution
------------

So how do we make this as painless to us and users as much as is possible?
The best way is to keep the surface area of the SPI or API small, create
solid boundaries, and avoid exposing implementation details and
implementation classes.

By reducing the surface area with implementation hiding we can effectively
limit exposure and reduce the probability of needing to make a change that
breaks with our user contract. You might be asking what's a real world
example of this for us in shared?

And incidentally this is one of the things I've been working on in my
branch.


Real World Example in Shared
--------------------------------------------

Let's take the o.a.d.s.ldap.message package as an example. This package
contains classes and interfaces modeling LDAP requests and responses: i.e.
AddRequest, DeleteResponse etc. It's in the shared-ldap module.

In this package, in addition to request response interfaces, we're exposing
implementation classes for them. The implementation classes, in turn have
dependencies on o.a.d.s.ldap.codec.* packages.
Not any more, I hope. We did a big refactoring last september in order to remove this coupling. Of course, we may have some remaining dependencies, but this is more or less not intentional.


Right not intentional and this is just one example in many. Look we've all used shared as a dumping ground. While our primary focus was a tough problem in Studio or in ApacheDS we put minimal energy into shared as we deposited some classes and interfaces into it. This is because the main focus was something else.

This is not to blame anyone. I am pointing out the problem, and pointing out a solution to it so we're not screwed by it. The web of dependencies in shared will f**k us down the line if we don't nix em now.
 

 This is because some
implementation classes depend on codec functionality which is an
implementation detail.

Not true anymore (or is it?).


Yeah there's some residual dependencies but not a big deal to fix. Trivial stuff. 

The work needed here is a joke really. The big issue with it is the impact the changes in shared have all over the place in Studio and ApacheDS and the fact that we're better off waiting for AP work to complete to merge.
 
This might be due to eager reuse or the addition of
utility methods into codec classes for convenience. Some of these
dependencies can be removed by breaking out non-implementation specific
methods and constants in codec classes into utility methods outside of the
package or the module all together. Furthermore the codec implementation
that handles [de]marshaling has to access package friendly (non-API) methods
on implementation classes while encoding.
Not sure that I get what you mean here. Can you be a bit more explicit ?

LdapEncoder accesses package friendly methods inside most message Impl clases to encode them. This also pulls into message dependencies from codec which can be hidden. But these are really easy to fix. We just need to know that the situation is there and get rid of it.
 

In the end, dependency upon further transitive dependencies are making us
expose almost all implementation classes in shared, and most can easily be
decoupled and hidden. It's effectively making everything in shared come
together in one big heap exposing way more than we want to.
It's quite impossible in Java to 'hide' all the classes that a user should not manipulate. Unless you use package protected classes, and it quickly has a limit, I would rather think in term of 'exposed' (ie documented) API.

OSGi bundles really helps in this respect. It fills in where Java left off. 

OSGi makes it so the (bundle) packaging coincides with module boundaries. In Java this is loose and there's leakage all over, as you say, it's very hard to hide all implementation classes.
 
That this documented API is gathered in one separate module for convenience is another aspect, but the user will still have to depend on all the other modules.


Certainly, you're right, dependencies will still exist. A codec will be depended upon for it's functionality even if we do hide the implementation details under the hood.

The value add here is not from avoiding a dependency. It's from not exposing more than we have to and being able to hide the implementation. This way we can change the implementation at will across point releases without having to bump up to a major revision.  
 
So all in all, should we define a module (a maven module) containing the public API and the associated implementation ? Probably (But this is not an absolute necessity). I guess this is what you have in mind, so let's see what's the proposal is...


We have multiple options for chopping this up. With bundles we have a nice tool to carve out physical not just logical boundaries to our API's and only expose those packages we need to show API users.
 

LDAP Client API
------------------------

Everyone agrees that this API is very important to get right with a 1.0.
Right now this API pulls in several public interfaces directly from shared.
Those interfaces also pull in some implementation classes. The logical API
extends into shared this way. Effectively the majority of shared is exposed
by the client API. The client API does not end at it's jar boundary.

All this exposure increases the chances of API change when all
implementation details are wide open and part of the client API.  And this
is what I'm trying to limit. There are ways we can decouple these
dependencies very nicely with a mixed bag of refactoring techniques while
breaking up shared-ldap into lesser more coherent modules. The idea is to
expose the bare minimum of only what we need to expose. Yes the shared code
has become very stable over time but the most stability is in the interfaces
and if we only expose these instead of implementation classes then we'll
have an awesome API that may remain 1.X for a while and not require
deprecations as new functionality is introduced.

How will you limit the visibility of the modules you don't want the user to be exposed to ?


A combination of refactoring techniques will be used to be able to better use standard Java protection mechanisms to hide implementation details combined with using OSGi bundles instead of Jars to only export those packages that we do want users to see.
 
Finishing Up the Example
-------------------------------------

So what concrete things can we do?

The biggest step is to hide as many of the implementation classes as
possible. In my experimental branch I started by:

    (1) Moving out methods and constants in codec classes causing
unnecessary dependencies from message package classes and interfaces. There
was a situation even where StringTools for example depended on codec
classes, and virtually everything doing string related operations used
StringTools there by causing man interdependencies. It then becomes a web of
dependencies across packages.
There is *one* method in StringTools that calls a codec method : Hex.encodeHex. It's a mistake, as we already have another StringTools method (toHexString) doing the same thing (to be double chekced). This is typically a wrong usage of a class from a wrong package, and we should get rid of such coupling.

Yep, no big deal something that gets fixed in seconds but we have a few of these kinds of examples. One by one they're nothing but all together they create this web making almost everything dependent transitively on each other. This is easy to fix.
 

This is extremely painful to do such a cleanup without first decoupling all the pieces by creating separate jars, before regrouping the packages back again.


Why bother regrouping? We can regroup things for convenience if people want a single jar without deps like the shard-all thingy.

We should not be uncomfortable having multiple modules to better decouple this big hunk of code, and isolate coherent pieces as units.
 
The question here is more to know how far we want to go, considering that shared contains 900 classes, more than 5600 methods and around 80 packages.


Yep it's big but the problem here is not massive. It starts slowly solving a couple things and once you decouple a few things, decoupling others becomes much much easier and a layout to all of it starts falling out nicely, which shows even if we dumped here and created some cleanup issues for ourselves the overall code really was written well.
 
    (2) Breaking up shared into multiple Maven modules so now there's the
following modules:

          o shared-util
          o asn1-api
          o asn1-ber
          o ldap-model
                 - name pkg
                 - message pkg (no impl classes)
                 - schema pkg
                 - cursor pkg
                 - filter pkg
                 - entry pkg
                 - constants pkg
          o ldap-codec (not complete)
I would not have 2 maven modules for asn1. It's probably overkilling. I would rather name the ldap-model ldap-api, because this is exactly what it is.

There are reasons for this to be able to get the codec to be separable. Once you get in and play with the little non-important details you'll probably come to the same conclusion yourself.

Let me propose a methodology we can follow here to speed things up without needlessly arguing each point because in the end we do in fact come to the same conclusions.

Let's just relax while decoupling about the number or name of modules. The first pass should be about breaking up dependencies to hide the implementation details so we're free to solve these aggressively without inhibitions.

Then once we see a clear dependency between modules, we can take another pass at consolidation as a separate concern and discussion. Until we see the real dependency picture fall out from refactoring it's moot over discussing it.
 
Otherwise, I like this decomposition.

There are a few more things we will have to discuss about :
- ldif (part of ldap-model/ldap-api)
- aci (but it may be in a separate module, a ADS specific one, as it's only good for ADS
- trigger (same as above)
- csn (maybe part of shared-util)
- dsml (a separate module ?)
- client api (connection, futures, exceptions) (part of ldap-model/ldap-api)
- i18n (separate module would be good)
- the schema loader probably deserves a separate ADS module too
- the schema converter too

The next step would be to make these artifacts into OSGi bundles. There will
be nothing special about it. I'm just going to leverage bundle packaging to
hide implementation classes which you cannot do as easily with regular jars
with explicit package exports.
That should be a no brainer.

Once this is done, we can export a minimal set of classes from the codec,
hide it's remainder, and have the model interfaces be the primary dependency
used by the client API without exposing implementation classes and keeping
the API weight (surface area) down.

There's a lot more to do, the job is 40% complete. The wait for the AP merge
makes this work feel moot since the merge is going to be nasty so I might
just redo this again after Emmanuel merges. That lets me be a bit more
agressive and experimental for now.
go for it. As soon as you have something stable, as it's all about moving pieces, we can do that bit by bit, instead of merging.

Plus if Pierre and Seelman decide to opt for using m2eclipse+Maven+Tyco (as
Jesse mentioned) for the Studio build then these refactorings a second time
will not incur manual fixing in Studio which depends on shared now. I can
refactor Studio at the same time.
The real issue here is m2eclipse : it's everything but usable for a project as big as ADS. I have tried it again one month ago, and it smell like Maven 1 to me...



Conclusions
-----------------

So this example shows some things we can do to make things tighter and
easier for us to better manage our API's. We can do anything we like to the
implementation to fix bugs and to improve performance in point releases
without impacting the minimal interfaces we expose for the API.

And it can be a good opportunity to clean up the shared module which has become a giant plate of spaghetti (with bolognese sauce).

We take similar steps inside the server to restrict down the exposed SPI
however using OSGi is probably not going to be an option there right away
since it gets more complicated. Here in shared I would use bundle packaging
just to hide implementation classes, not to define services etc.
Also there are some classes that were proposed for shared, i.e. DnNode which
at this point in time are specific to the server. Sure Studio might use
these classes eventually, however these classes are not generic LDAP. These
classes can stay in shared but they should be kept in a module separate from
the ldap-model for example.
Agreed. There may be other classes to, they have to be identified.

Why you may ask? Because these classes are not
generic LDAP classes (like Entry, or Dn, or Cursor is generic and) are not
needed by every client, nor are they viable for every server a client
connects to. They only serve a purpose when used in Studio, connecting to
ApacheDS.
They are helper classes. They certainly don't belongs to ldap-model/ldap-api, and if they have to stay in shared, I would like to move it to utils.


We discussed this last night. Just wanted to point out the clarification you made to me. 

Shared will have shared-utils for generic utility classes that can be used by anything not just LDAP code. Then there may be a shared-ldap-util but I think this might be overkill and not such a good idea: let me explain technically why:

If we dump these ldap specific utility classes into an ldap-util, then a dependency to one util class pulls in utility classes in the rest of the module increasing footprint perhaps needlessly.

What is needed for minimal generic client operation should be kept together with as little dependencies as possible. No need to expose the plethora of utility classes we have amassed in there. 

Yes they are very useful utilities but it's not about packaging freebees it's about keeping things small, tight, and minimizing exposure. We can package these things into a separate jar and only use them in studio and in apacheds.

So in conclusion what I am saying is the general formula we have become accustomed to where we throw all utility classes into one module no longer works blindly for us. We need to think about what needless dependencies and classes this is including. 
 
DnNode might be needed by Studio in the future for making a plugin and
widget that allows users to graphically manage the boundaries of
administrative areas, however it's not something every client needs, and it
certainly is not something needed by a generic client connecting to every
server.
--> utils.


Again let's be more specific but let's not overly force ourselves right away - perhaps there's a better name for this trapped in our heads. Give it time to trickle out. As we decouple and break things appart this will become much easier to see clearly. Just a general MO suggestion.

--
Alex Karasulu
My Blog :: http://www.jroller.com/akarasulu/
Apache Directory Server :: http://directory.apache.org
Apache MINA :: http://mina.apache.org
To set up a meeting with me: http://tungle.me/AlexKarasulu