Nino,
You are absolutely right that there are drawbacks with the single file
storage approach. I'll try to outline why we decided to go that
direction, as our original design actually used multiple files. Most
of the issues we’ve seen when multiple developers make changes to an
application can be attributed to two primary causes: the information
is stored in XML format and there is a separation between the storage
format and the medium familiar to the developer. Let’s take a look a
closer look at these problems.
The XML format is fabulous for certain use cases. Storing highly
structured data that contains large amounts of text that needs to be
accessible on multiple platforms and manageable by humans is one of
them. And this is exactly the type of data that makes up our
application definitions. The problem is that is gives the diff engine
of most source control systems fits. Nearly by definition, all data
markers in XML are variable length in both individual composition
(attributes) or number and order (sub elements). In my experience the
most usual culprits of incorrect difference calculation are elements
with multiple attributes, especially when the number or which
attributes are dynamic, and the end tags of elements with children. To
help avoid the havoc caused by these instances, the application
document generated by the VTP follows some strict rules:
* XML indentation is fully enforced on all elements and levels of
nesting
* Start tag pre-ambles are always placed on their own line
* Element attributes are always placed on their own line
* Element attributes always occur in the same order regardless of
presence
* Element end tags are always placed on their own line
* Element collections are always written in the same order they
were read or created unless there is an overriding ordering
mechanism
However, even with all these rules in place, there are still instances
of conflict when merging changes from independent sources. This brings
us to the other basic driver; the developer is not familiar with the
XML format. Our development environment presents the developer with a
100% graphical representation of all information present in the XML
document. This goes so far as the editor doesn’t even bother with the
underlying file until a change needs to be committed to storage. Often
times, the developer doesn’t even have experience with XML or other
central technologies employed by the VTP let alone a working knowledge
of the schema and nuances of our specific XML language. Eclipse is
great in that it happily displays the head to head conflict resolution
UI. Given the target audience, this can result in even further data
loss depending on their tendency to click buttons.
I’ve now done something I hate to do, describe an array of problems
without really offering any solutions. I’m not sure if there are any
“good and easy” ones. OK, back to why we migrated to single file over
multiple. The main reason was that using multiple files to store the
data opened our project structure up to a possible inconsistent state,
depending on user behavior. In a traditional java project, having many
different files with changes in them and various states of committal
to the source repository is normal and often doesn’t even result in an
inconsistent state. However, our applications are very integrated and
represent a continuous flow of logic from the single entry point to
its final states. There are very few edit oriented operations that
could be performed that wouldn’t leave the entire project in an
inconsistent state if all documents weren’t checked in. Although this
type of situation is fairly common for those of us who are familiar
with the standard software development lifecycle, it can be quite a
learning curve for much of the target audience.
Another reason we decided to perform the migration was all the other
issues with XML were still present in the multiple file approach, so
we decided to merge all the files into a single source. To be honest,
I’m not entirely sold on this approach, but it solved more issues than
it presented at the time. Choices like these really resonate with what
Kent Beck discussed in his /Design is an Island/ piece on Planet
Eclipse. I would love to enable a greater liberty in changes being
made by multiple developers. I would like to mention that with a
moderate amount of discipline, multiple developers can still edit the
same project. If the developers remain within say a single canvas or
don’t modify the same canvas, the likelihood of a conflict is greatly
reduced thanks to the efforts I mentioned above. Also, there is the
ability to separate your application into several independent projects
as application fragments. These can obviously be edited without fear
of collision.
Ultimately, I think the only real solution is to provide a completely
custom conflict resolution UI that allows the application designer to
merge changes using the medium they are used to. Showing the changes
in a head-to-head visual display and simple point and click change
acceptance. This would almost be a project in and of itself and
consequently has been a lower priority than other more easily
reachable fruit. I’ve been recently playing around with a couple of
ideas on how to make the XML rules we follow more effective. My
favorite one so far is to sort of watermark the end element tag so as
to identify it uniquely even within a list of similar elements. For
example, most of our elements represent objects that have an
identifying attribute like ‘id’. The line containing the end tag would
have an XML comment added just after the tag content that contained
the identifier. This would allow the line based diff engines to more
easily match groups of lines together in cohesive segments. I haven’t
fully worked out the details yet.
Trip Gilman
On 4/14/09 10:57 AM, "Nino Martinez" <nino.martinez.wael@xxxxxxxxx> wrote:
> Hi Trip
>
> Now that you mention a thing about the runtime platform, as I see it the
> interactions you describe in the visual layout all goes into one xml
> container? Im wondering why it's done this way? Nortel's SCE, which im
> heading into using everyday, also has this approach.
>
> Making it impossible to be more than one dev on projects. Since you
> overlap changes in the xml file.. Or am I wrong about VTP (I hope so)..
>
> If it instead could be contained in some code generation or something
> like it, maybe taking advantage of java7's scripting possibilities (when
> it comes) or something else to leverage this problem.. However this
> would probably impose a big code change..
>
> regards Nino
>
>
> On Sun, 2009-04-12 at 19:27 -0500, Trip Gilman wrote:
>> David
>>
>> First I'd like to thank you for your interest in the VTP and welcome you
>> back to the industry. I've provided a broad overview of where we are
now as
>> a project and also weaved some of my detailed thoughts on your questions
>> into your original text below.
>>
>> You are correct in that the current VTP is a complete departure from the
>> codebase originally contributed by IBM. I would like to take a moment to
>> point out some of the fundamental changes between the original
vision and
>> our current direction and then discuss the reasoning for our approach.
>>
>> First, the xml editors for GRXML grammars have been replaced with the
>> generic xml editor provided by the basic eclipse development
framework. The
>> original VTP utilized a customized xml editor based on those
provided by the
>> Web Tools Project. The removal of this dependency reduced the
overall size
>> of our software by nearly 100M and eliminated several layers of
complexity
>> for no real loss of functionality.
>>
>> Second, and probably the most centrally important, the VTP employs a
common
>> runtime architecture as opposed to code generation. I don't want to
focus
>> on the choice between the two, but I would like to illustrate how we
>> leverage our choice of a runtime architecture.
>>
>> The voice tools project has almost become a misnomer. Our
development and
>> runtime frameworks are no longer just for VXML, but have evolved into a
>> generic platform for describing interactions. The core of the
development
>> environment is blissfully unaware that the user intends to use the
>> application they are creating as the source of a VXML interaction. The
>> runtime environment executes the interaction within a process engine
that
>> emits basic interactive commands that are interpreted into VXML at
the final
>> moment. This is important to understand as this philosophy is pervasive
>> throughout the codebase.
>>
>> The user creates applications by adding Actions to a canvas and then
>> defining the transitions between those actions. The most common Actions,
>> such as a Menu or a Prompt(Output), are standard components provided
by the
>> framework. However, third-party developers can easily and quickly create
>> their own action libraries for users to use. I want to reiterate that
>> although these actions may share the names of the elements that make up
>> VXML, they are not the same.
>>
>> The designer has an abstraction framework that models the nuances of
>> interaction types on these actions, allowing interaction specific
>> configuration. For example, the VTP comes with an implementation of the
>> Voice (VXML) interaction type. This provides interaction specific UIs to
>> manage VXML settings such as barge-in and dtmf/speech entry.
>>
>> This framework has been used to create other interaction types such as
>> Instant Messaging. An application can support any combination of
>> interaction types the user has available to them. This is very
powerful in
>> that the same application deployed once, could drive the interactions of
>> both telephone and IM based customers.
>>
>> On the runtime side, things are equally abstracted. As I mentioned
earlier,
>> the artifacts produced by our designer (upon export) are actually
process
>> meta-language documents. These documents are executed by the process
engine
>> at the core of the runtime. As the process is executed, interactive
>> commands are generated. These commands are then filtered through a
platform
>> abstraction layer and evaluated. The VTP includes a platform layer
designed
>> for VXML interactions that is capable of transforming the interactive
>> commands into their representative VXML documents.
>>
>> The platform abstraction layer is powerful on two fronts. First, it
allows
>> the VXML generation process to be customized to the specific VXML
Browser
>> platform requesting the document. This means that a single application
>> instance can service multiple VXML platforms simultaneously without
>> re-export or compilation. The VXML session could even bounce back
and forth
>> between the platforms without issue.
>>
>> Second, the platform abstraction is open to third-parties to add
>> implementations for additional platform types and output mechanisms. The
>> VXML platform generates XML documents and processes requests and
responses
>> over HTTP. Other implementations could simply execute direct API
calls to
>> other software or hardware, never producing a document artifact.
They could
>> also support transports other than HTTP.
>>
>> I hope this helps illuminate the project for you.
>>
>> Trip Gilman
>>
>>
>> On 4/10/09 8:33 PM, "David Reich" <David.Reich@xxxxxxxxxxx> wrote:
>>
>>> Hi,
>>>
>>> First, let me (re)introduce myself. David Reich, formerly of IBM. I was
>>> involved in the creation of VoiceXML and was the guy who was the
initial
>>> driver behind all of the Eclipse-based voice tooling, first from IBM's
>>> toolkit, and it's subsequent donation to Eclipse and VTP. I've
since left
>>> IBM
>>> and as you know, IBM's involvement in voice has waned. But, I am
now with a
>>> new company, happily back in VoiceXML, CCXML and SRGS, and we are
doing a
>>> lot
>>> with VoiceXML and CCXML and I was brought in to drive our
development team,
>>> and one of the areas I am looking at is the development of our
application
>>> artifacts, and VoiceXML application generation.
>>>
>>> With that, I've spen the last couple of days figuring out where
things are
>>> with this project, how we can use it, how we might give back to it
and so
>>> on.
>>> From what I've been able to see thus far, it looks like the current
code
>>> base
>>> bears little to no resemblance to the VTP 1.0 code that came in
from my
>>> former
>>> team, but, on the other hand, there is a nicely-coming-along visual
editor
>>> without which all VTP would be is the XML editor with DTDs plugged
in and
>>> while nice, not really compelling.
>>>
>>> I was hoping to provoke some discussion and see about some
combination of
>>> code
>>> to create an even more gener(ic)ally useful environment. With that,
some
>>> thoughts and the hope we can put some other things together. In no
specific
>>> order:
>>>
>>>
>>> * The original IBM code seems to be gone. When things moved to SVN, the
>>> CVS repository went away, so unless someone has that source code,
it's gone
>>> forever (I've checked with the (remnants of) the IBM team, and
without some
>>> serious effort, it's not readily available. So, how can we get that
back,
>>> even if we only want to take small bits (or even none?) of it?
There's got
>>> to
>>> be useful things there, and we should still look at that.
>>
>> The original IBM source still exists, however it is not readily
available as
>> it has been archived from CVS by the webmasters to reclaim space and
keep
>> the repositories current. If you would like a copy of the archive, I
would
>> be more than happy to make the request for you.
>>
>>> * In the original tools, of which the VTP 1.0 code was a part, the
vision
>>> was a call flow builder similar to what WebMethods has been working
on, but
>>> rather than generate a metalanguage, have it generate VoiceXML and
have a
>>> tabbed view (kinda like FrontPage) where you see the call flow
view, and the
>>> VoiceXML source view, and the editor would allow you to tweak the
VoiceXML
>>> in
>>> source edit mode, affording finer-grained control of the VoiceXML.
>>
>> Very early on we abandoned code generation in favor of a runtime
>> architecture to overcome several challenges we were faced with when
>> developing applications in the field. Each time we encountered a new
VXML
>> platform, there were always a myriad of minute differences or
interesting
>> interpretations of the specification. We considered just coding these
>> changes into the code generation piece until we ran into a client
that had
>> multiple platforms in service. The solution was to export the
application
>> twice, once for each platform. This instance was one of the driving
forces
>> in the transition to a runtime architecture.
>>
>> Another driving factor of the move, was the realization that we were
already
>> building a runtime, just not officially and not effectively. We had
>> developed a loose library of utility functions and were using them
over and
>> over again in all our applications. More and more functionality was
being
>> encoded into this framework in an effort to hasten development.
Making the
>> runtime a first class citizen in the architecture allowed us to
leverage the
>> advances we had made while removing most of the hurdled we were facing.
>>
>>> * Linkages to the Eclipse framework for task-list items in validation,
>>> warnings, errors, etc.
>>
>> Absolutely, integration to the tagging systems in Eclipse is
currently on
>> the wish list.
>>
>>> * Enabling (as Randy has told me is the direction) a derivative or
>>> extension plugin to generate code that spews forth VoiceXML. This
could
>>> take
>>> the form of the OpenMethods servlet, or even perhaps an extension
one could
>>> write
>>>
>>> to do JSPs, or other types of markup/tag/metalanguage since not
everyone
>>> will
>>> want just the VoiceXML from the OpenMethods servlet, or wants to make
>>> further
>>> tweaks to the VoiceXML.
>>
>> I won't say that it would be impossible to generate the resulting
VXML from
>> an application statically, but at some point you'd end up recreating the
>> logic inherent in the runtime in JSP or some other language. In the
case of
>> multi-modal applications, simply duplicate the application several times
>> over.
>>
>> On a side note, the modification of generated code after export can be a
>> very tempting practice as many problems can be solved quickly this way.
>> However, I never recommend the use of this. It places a process
requirement
>> on the development and deployment staff to always re-apply the
changes to
>> the code each time there is an update. What happens if changes made
to the
>> code require the workaround to be updated? This can easily balloon
into a
>> maintenance nightmare. I would say that it should be the goal of any
>> system, whether code generation or runtime, to avoid the need of
this at all
>> costs.
>>
>>> * A broader property sheet for each call element in the palette on the
>>> canvas to specify items such as barge-in, SSML tags or information,
prompt
>>> audio files to be played and so on. Specifically, the block would
play one
>>> of
>>> a set of prompts based on current state (such as "Welcome to..." or
>>> "Welcome
>>> back to..." based on some global variables in the ECMAscript for
example.
>>> This tool should should hopefully allow for the export of this
property
>>> sheet
>>> data in different formats, but then again, as open source, we can work
>>> together on extensions for our specific needs and write the hooks and
>>> extension code for different needs.
>>
>> The designer already supports this type of configuration in several
ways.
>> At a base level, the developer can use Javascript variables to indicate
>> prompt content. SSML can also be included in prompt content simply
as text.
>> At a higher level, the designer employs a system called branding to
allow
>> very detailed configuration for different user segments. Branding can be
>> used for personalization, line of business differentiation, or even
>> multi-tenanting of an application.
>>
>> On the export aspect of this, I assume you intend to use the artifacts
>> within a static JSP/VXML document.
>>
>>> * Have the palette be extensible where one can take prebuilt
modules (say
>>> an authentication subflow) add them to the palette, and drop it
into a new
>>> or
>>> existing application.
>>
>> The designer is fully extensible. Third-parties can easily publish
>> additional modules for use. Developers are also able to package up
common
>> call flow elements into reusable components called dialogs or even
>> significant portions of a call flow into a fragment.
>>
>>> That's enough or now, but you get the general idea. I'm hoping we
can make
>>> this a bit more granular and extensible, and some of that is in the
earlier
>>> code (yes, I also have a fondness for my baby ;-) and marrying that
with the
>>> newer code, I think we can do a lot here.
>>>
>>> Comments?
>>>
>>> Thanks....
>>>
>>> David Reich
>>> AIT Architect, Adeptra, Inc.
>>>
>>>
>>>
>>> _______________________________________________
>>> vtp-dev mailing list
>>> vtp-dev@xxxxxxxxxxx
>>> https://dev.eclipse.org/mailman/listinfo/vtp-dev
>>
>> _______________________________________________
>> vtp-dev mailing list
>> vtp-dev@xxxxxxxxxxx
>> https://dev.eclipse.org/mailman/listinfo/vtp-dev
>
> _______________________________________________
> vtp-dev mailing list
> vtp-dev@xxxxxxxxxxx
> https://dev.eclipse.org/mailman/listinfo/vtp-dev
------------------------------------------------------------------------
_______________________________________________
vtp-dev mailing list
vtp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/vtp-dev