Data Services Connection

Why Data Modeling is the Wrong Tool for Integrating Data Models

noreply@blogger.com (Eric Samson) — Thu, 07 May 2009 20:50:00 +0000

Fresh ideas in this article about how to build common data models, even if I wouldn't agree with everything.

Tactical Data Integration with JBoss Federate

noreply@blogger.com (Eric Samson) — Thu, 07 May 2009 20:48:00 +0000

Video about Federate, the new open source version of the former MetaMatrix EII product:

http://architects.dzone.com/videos/screencast-tactical-data.

Federate Project. (most of links don't work)

DryadLINQ

noreply@blogger.com (Eric Samson) — Thu, 07 May 2009 20:44:00 +0000

Quite interesting automatisation of distributed queries:

http://www.infoq.com/news/2009/05/DryadLINQ

Microsoft research project.

Data Services presentation at The Open Group Conference

noreply@blogger.com (Eric Samson) — Tue, 03 Feb 2009 21:24:00 +0000

I did a presentation about Data Services yesterday at the Open Group Conference in San Diego. The session went very well and we had a series of good questions / feedback:

How is security managed in a DSP?
What about transaction management?
Is that a good platform for mobile applications?
Where are metadata deployed?
Can we call this a Data Hub?
What about deployment of a DSP within an ESB?
What are the real infrastructure costs of WOA and how to compute/compare them?
What is the link between a DSP and MDM? (see my recent article about this subject)

Seems the benefits of a Data Services Platform are now widely accepted.

Pseudo transactions with non-XA resources

noreply@blogger.com (Eric Samson) — Tue, 03 Feb 2009 20:48:00 +0000

I've seen this article about Pseudo-transactions with non XA-resources on TheServerSide. Well, basically what they do is a manual compensation of the non-transactional resources. There are many ways to implement this, some of them could be quite operational, but this is not my point today.

It is quite possible in the future that we'll have to think differently about transactions. First of all, because it will be always difficult to manage non-transactional resources. Second, the real problems will raise when we'll start to deal with asynchronous data access patterns, e.g. when you call a data services and you don't wait for the answer, you'll get it later (maybe).

The multi-agents systems and the actor languages give interesting ideas about this. An actor language is like an object language in which every method call is asynchronous by default, any required synchronisation having to be explicitely defined and managed (continuations, etc.).

I've recently read interesting articles about this. Some people are saying we can just drop the whole notion of transactions. It could be shocking at the beginning but they have good arguments. Other ones think we must change the problem until we can have atomic transactions: see for instance "Life beyond distributed transactions" by Pat Helland (from Amazon) or the Saga transaction system from Arnon Rotem.

The Role of a Data Services Layer in SOA

noreply@blogger.com (Eric Samson) — Mon, 19 Jan 2009 19:49:00 +0000

UK-trade journal, Database & Network Journal, has published one of my articles in their January issue: “Bringing Data into Service: The Role of a Data Services Layer in SOA”. I'm not aware of any Web version of it, I'm looking for a solution to make it available online. Stay tuned.

You can read the article here.

How MDM and Data Services can be Complementary

noreply@blogger.com (Eric Samson) — Mon, 19 Jan 2009 19:41:00 +0000

I'm pleased to inform you that DM Review has published my article about Data Services Platforms and Master Data Management.

In the past, I used to think that MDM was mostly an issue for EII products. Today, most EII have turned themselves into Data Services, but there are other approaches to Data Services Platforms. Based on my discussions my some MDM vendors I'm now convinced that DSP platforms have an important role to play in that space. You'll see all the details in the article.

SOA is dead?

noreply@blogger.com (Eric Samson) — Thu, 08 Jan 2009 16:40:00 +0000

First of all, I wish you all the best for 2009!

This new year starts with an obituary news, from Ann Thomas Mann (Burton Group): SOA is dead. This blog entry has generated a lot of reactions, and that was probably one of its aims:

(just to mention a few of them).

In fact, what she says is that service-orientation is still needed but the SOA TLA is now too heavily associated with the idea of project failure.

First comments:

Seems to me all this is more provocation than anything else, done to generate some noise and visibility. Regarding this, it is a success.
She is just writing what many were thinking and even saying since many years.
People tend to quickly burn the idols they have created themselves. So that they can quickly idolize another one... with the same unreasonable passion.
Sometimes it needs time, even for a good idea, to impose itself. Unfortunately, the World and the software industry are now stuck in short-term visions and immediate ROI, there is no more time for ideas to mature. New concepts now have to be perfect from inception, there is no room for improvement.
If we agree that SOA is not a technology but rather a way of thinking software architectures, then saying that "SOA is dead but there is still a need for the idea of services" is a kind of oxymore.
If SOA is really wrong, changing it's name won't fix it. WOA, RESTFul, Mashups won't save services if the main idea behind it is not good. The problem cannot just be with the SOA name.

So, what was really wrong with this supposedly dying SOA?

SOA is much more complicated than expected. Remember: the 'S' in SOAP means "simple". SOA was supposed to be simpler than CORBA and J2E. But why SOA would have been less complicated than its predecessors? When you want to design non-trivial business applications by assembling distributed components you will face complexity. Using HTTP / SOAP instead of IIOP and WSDL instead of IDL will not fundamentally change the thing. The problem is still complex and unfortunately (or not) we don't have easy solutions to complex problems yet. Thinking large systems as communicating distributed agents is not so easy. It is all about parallel asynchronous programming. Most developers are not trained for that, as they were not trained to be Java gurus. It is so funny to see that today SOA is attacked on the simplicity front.
SOA is too much process-centric. And once again, it was the same with CORBA and J2E. All distributed arhcitectures are first trying to solve the easy problem first: distribute processes. Then they start to deal with data and the real problems raise.

But it is not necessarily too late for SOA, as far as we can reconcile data and processes, states and behaviors, attributes and methods, however you name it.

So let's go back to the fundamentals. David Linthicum is thinking in the right direction: How to deal with data in SOA?

That's where we are today in SOA. Data Services are clearly going in the right direction. Customers will be able to solve some of their data integration issues with SOA, and they will have time to understand the value of this approach. To some extent, Data Services are at the cross between the promises of SOA and OO: states encapsulated by business methods.

Obviously, we'll need to go further, move from simple data federation to adaptative mediation, but we'll have time as we are solving real problems in the meantime.

Data Services should not be too much focused on names like SOA, WOA, RESTFul, mashups. New architectures are not replacing old ones (even if their proponents always tend to think it, at the beginning). They just complement them. Data Services platforms should equally support all these styles. Data Services Platforms should publish integrated data as SOA-or-not services, but also as POJOs, xDBC... The same way they have to connect to SOA-or-not data services as data sources, but also to RDBMS, packaged applications, mainframe transactions and screens, etc. The real IT world will certainly never be all-SOA, but equally it won't be all-non-SOA.

Services are basically components. There are many deep reasons why imposing a component-oriented design has continuously failed for decades. SOA is a new try and it won't be the last one. Component-based progamming models have failed for cultural reasons (programmers are not ready for that) and also because of the Lego model. You can't build a complex system by manually linking thousands of atomic components. SOA is way too much manually hard-coded today. We need to be able to dynamically generate, publish and compose services at runtime. We need to design new data access patterns, maybe based on asynchronous, reactive models. This will require advanced metadata, more semantic metadata and dynamic computation of chains of service calls at runtime. All this won't be achieved in few years.

In the meantime, programmers will progressively have to accept they won't control the flow of service calls, exactly as DBAs had to accept they won't tune all SQL statements any longer with the raise of persistence technologies. Each time you add abstraction layers you are able to do more, but conversely you lose some visibility and control on the lower layers.

So, the real debate is not what's next after SOA?, or which style of service-orientation will win? REST is not simpler than SOA if you want to manage complex things by manually connecting services. If one day we want to solve complex problems with simple solutions, then we'll need new abstraction layers, we need dynamic SOA. And for that, we need to stop manually calling and connecting services. Regarding this SOA versus WOA or SOAP versus REST is not the point at all. Services are a required piece in the puzzle, but we need new abstractions on top of them.

See the funny REST is dead blog entry here.

EMC Atmos

noreply@blogger.com (Eric Samson) — Mon, 15 Dec 2008 17:45:00 +0000

Let's try to be more positive about new data sources for the cloud, after the harsh comment of a reader seen in my previous post. Last week, I had the opportunity to discover Atmos, a new Cloud Optimized Storage (COS) from EMC.

What I particularly liked is the fact it is based on an object-oriented distributed file system. Microsoft wanted to build such a file system few years ago and eventually they gave up. EMC did it. I still need to dig into the details of that product, but I really think the possibilities of a true OOFS are almost impossible to imagine right now.

Performance of new databases

noreply@blogger.com (Eric Samson) — Mon, 15 Dec 2008 17:17:00 +0000

Interesting performance comparison from Oakleaf. They compare an Amazon EC2 application deployed with different storage options (SQL Server, SimpleDB and Azure Table Service).

I reproduce the result table, if you have no time to read the article:

A reader's comment:

A major piece of work by OakLeaf … has confirmed my suspicions about how inappropriate all the cloud name/value entity store technologies are for serious SaaS developers.
The Google AppEngine Datastore, Amazon’s SimpleDB and Windows Azure have chronic performance problems relative to conventional database throughput. Ultimately, the inherent inefficiencies of these storage options will hit hourly cloud renters in the pocket.

Semantic SOA

noreply@blogger.com (Eric Samson) — Mon, 15 Dec 2008 16:52:00 +0000

Seen this post on InfoQ, OASIS has released a new version of their Reference Model for SOA.

It is well admitted that if we want to have SOA more policy-driven we must stop to manually compose services together. And this can mostly be achieved through extended metadata (I mean more than WSDL). WSDL was a nice beginning but:

Is is limited to Web services, and not all services are Web services. For instance some existing mainframes applications, transactions, green screens, APIs of packaged applications and Stateless Session EJBs are also services.
Relationships between entities are missing.
There is nothing about behavior of services.

The whole Semantic Web is an ambitious set of projects. However, when it comes to data access in SOA, we have to deal with the same kind of issues. Accessing a data services layer is more complicated than accessing a databaase, because it publishes a business interface instead of a technical interface. Therefore, if we want to dynamically compose data services together, as opposed to hard-code the combinations, we need some kind of semantic metadata about the data manipulation behavior.

That dynamic aspect is fundamental for modern Data Services Platforms and is quite often missing in first generation technologies.

The first step of SOA was all about designing, deploying and consuming services. The second step is now all about dynamically designing, dynamically deploying and dynamically consuming services. We cannot spend all our time in manually connecting thousands of services together.

The Information Perspective of SOA Design

noreply@blogger.com (Eric Samson) — Mon, 15 Dec 2008 16:07:00 +0000

Seen this post on InfoQ. I like to see that most people agree on the need to reintroduce the data aspect into SOA. You should read the full article from IBM, but they basically insist on 3 points:

Define data semantics.
Canonical modeling.
Data quality.

The 2 first points are at the heart of any Data Services Platform.

Defining Data Semantics is very important and should be done with extended metadata. The richer the metadata are, the further we can decouple data sources form data consumers. And this is a strong requirement if you want to be more "policy-driven" and less hard-coded in your data access strategies.

Canonical modeling is sometimes seen as a burden by database purists. However, it is required as soon as you need to federate heterogeneous data sources, that is the common case in SOA. On top of that, canonical models offer a more business-friendly view of data. The question remain about the scope of canonical modeling. Very large business models, at the enterprise level, are known to be very difficult to design and to maintain. DSP must offer a way to select the best granularity for canonical models. One model for all applications is a myth (at least today), but conversely, one model per application does not allow reaching the promises of reuse of SOA. On this aspect it is also interesting to track what vertical standardization efforts (SID, HL7, Acord, SWIFT, etc.) could bring on the table. We now see users starting to use the Telco SID model outside of the Telco market, for instance (from a 40,000 ft perspective, a customer is a customer).

The third one, is a market by itself (MDM) but should also be accessible from within a Data Services Platform. Relationhips between MDM and DSP technologies are multiple. DSP can be used as a synchronization layer for MDM products. DSP can support reference data, as a new kind of data sources, capturing their specific meaning (reference data can be a link to the real data, of maintain a link with it). DSP could use data cleaning services to improve data quality. These are just examples.

All in all, this is all going in the right direction.

Criticism of Java persistence

noreply@blogger.com (Eric Samson) — Mon, 08 Dec 2008 14:06:00 +0000

The recent Criticsm of Java Persistence thread on TheServerSide shows us once again that data access is still a very sensitive, emotional and almost religious topic.
I have said what I had to say in the thread itself, so no need to duplicate this here.

Many people have wrong ideas about persistence in general, and many people have their own views about how to manage it (while conversely very few people are publicly defending their opinion about how to manage pages in an operating system for instance). There is definitely something special about data access and persistence.

NB: I really think ODBMS vendors should stop communicating on the "the best mapping is no mapping" moto. First it is not true, and second it does not help them.

ODMG's not dead?

noreply@blogger.com (Eric Samson) — Fri, 05 Dec 2008 16:21:00 +0000

Seems OMG will host an Object Database Standard Definition Scope meeting in Santa Clara, next week.

Versant acquired db4o yesterday

noreply@blogger.com (Eric Samson) — Fri, 05 Dec 2008 15:43:00 +0000

See the news from the Versant site and from the db4o blog.

Well, that the Enterprise ODBMS buying the embedded ODBMS. Don't know what it means to the ODBMS community. Does it means that eventually the db4o business model didn't work as expected despite the good image of the product and company on the market? I also don't clearly see what it means to the Poet's FastObject part of Versant...

I know both products quite well, and I know both teams share some common genuine values around quality and performance. Even if I now work for a company who owns ObjectStore, I wanted to send a sincere "good luck" to them.

LINK like initiatives for Java

noreply@blogger.com (Eric Samson) — Tue, 02 Dec 2008 16:35:00 +0000

LiquidForm: http://code.google.com/p/liquidform/
JaQu from the H2 database.
db4o is working on some kind of LINQ equivalent I think.
JEQUEL.
Quaere.
LINQ for Java (Google Group).
Query DSL.

OJM

noreply@blogger.com (Eric Samson) — Wed, 26 Nov 2008 19:37:00 +0000

Object to JDBC Mapping: OJM.
Seems to me when I first saw that new TLA that it was a pure oxymore.

Reading the CPO web site just confirmed this. There are so many strange amd sometimes inaccurate statements on this page. The author seems to ignore what modern persistence is.
I really hate to be harsh like that, but is there really a need in 2008, for yet another JDBC abstraction?

Different problems, different databases

noreply@blogger.com (Eric Samson) — Wed, 26 Nov 2008 18:43:00 +0000

The big promise of relational databases was to have a unique, single technology for all our data storage needs. The main idea was to separate data from applications manipulating data.
Having data models too much coupled with applications model has indeed been recognized in the 70s as one of the main problems preventing IT flexibility (for instance, because producing new reports for business users required to go back to a full development cycle).

But at the same time, the design of software made significant progress by recommending encapsulating data (state) within methods (behavior), clearly going in an opposite direction. This all created a lot of stress and noises in the software industry, and eventually the emergence of persistence technologies.

But what really means decoupling data from applications? It mostly consists in removing explicit directional relationships from database schemas, so that data views can later be recombined in any way. When you think about it, it just means that relationships where poorly represented in programming languages, and it is still true with most modern languages including Java and C#. But to be honest, relationships are also poorly represented in relational models and this is not the nasty foreign keys that will change anything to it.

The fact is that the next real big revolution in the IT world would be a first comprehensive support for the notion of relationships.

Object database vendors failed to impose ODBMS while it was the most relevant choice for Java, at least from the technical point of view. There are many reasons to that:

Some first ODBMS implentations were very bad in terms of database administration, ad-hoc queries and overall performance. They were not really database but rather a storage mechanism for in-memory object pages.
The "ecology" never really started arounbd ODBMS (reporting tool...).
ODBMS started in the late 80s, exactly when RDBMS were just about to gain mometum on the market, it was not the right time to impose a new database technology.
Then major ODBMS vendors raised money from IPOs in the 95, in a very quiet time with no opportunity for expansion and no real need for money, therefore most of that money has been waste for nothing.
In 1998, the ODBMS vendors missed the Internet wave, mostly because of the XML mania at that time.
Then they surrendered and tried to reposition themselves as cache (Versant, Gemstone) or XML storage (Objectstore).

Hopefully, the XML database market never really emerged, despite the huge XML hype, probably because everybody understands XML is a good exchange format but a very bad (too verbose) storage format. The big problem with XML is still that it tends to impose hierarchical models, which to some extents are a kind of regression in our industry.

You can easily have an object or XML layer on top of any kind of storage, including relational (see IBM pureXML, for instance). Probably the best approach would be to have neutral and efficient storages, with multiple interfaces around them. It could be a kind of relational storage with the notion of relationship efficiently supported.

Internet, SOA, WOA, Web 2.0, mashups, etc. all favor a style where business functionalities become independent, and thus will have their own storage (as they cannot share a common database any longer, as they are really physically distributed).

It now seems some vendors are now trying to push the idea that even this low-level underlying storage layer, should have different foundations, depending on the kind of problems addressed. That's why the vertical model (storage primarily organized by columns instead of rows, like in Vertica) or the key-value model are quickly growing these days.

They will certainly not replace RDBMS systems any soon, as some are already claiming, but they will maybe impose themselves in some situations. Seems to me the time of the omniscient relational model is about to decline, even if it will remain present for decades.

Data services direclty impacts the database world because:

We have more and more data sources to access from a even a simple business application.
The notion of transaction is change.
We more and more frequently have to support asynchronous data access.
It becomes not only possible but also mandatory to access any kind of data sources, not only relational ones.
Databases are progressively commoditized, and their advanced features will move to intermediate mediation layers.
Then it is possible to choose the best database technology for a given need, at any time.

New databases

noreply@blogger.com (Eric Samson) — Wed, 26 Nov 2008 18:19:00 +0000

Following my previous blog about the future of databases, I've seen these recent posts on TheServerSide:

Interview on CouchDB, the Apache project for a document database, written in Erlang, with an HTTP/REST/JSON API, already mentioned in this blog. Interesting point of view at the end about BigTable. For sure a big scalable, persistent Map is certainly interesting in some very specific cases, but it is not really a database at the end.
Scalaris, a scalable, transactional data store for Web 2.0 services. Yet another distributed, persistent key-value system. That one shares many design ideas with BigTable, while, like CouchDB, it is written in Erlang. The big addition seems to be a better support for transactions. See OnScale for more information (videos, slides...)

The Future of Databases

noreply@blogger.com (Eric Samson) — Tue, 25 Nov 2008 15:29:00 +0000

Last week in San Jose, during the Data Services World event, I've been participating to several discussions (private journalists briefings, public panels) about the Future of Databases in the Cloud Computing. There are people who really seriously think today that RDBMS will soon disappear because of the Cloud.

I have seen that the same topic has also been discussed at various locations at the same time, including interesting point of view from Martin Fowler.

New technologies for databases can be roughly divided into two groups:

New kind of database technologies, tuned by design for the Cloud like Vertica, CouchDB, SimpleDB and other products alike that I've already mentioned in that blog.
New deployment and access for RDBMS, known as Database-as-a-Service. Basically, the database is remotely hosted and administrate, but you still access it through SQL over HTTP or SOAP/REST.

Having worked in the past for an ODBMS vendor I know how difficult it is to convince CIOs, project managers, architects, developers and DBAs to move from RDBMS. There is a kind of religion about relational theory. My take is that RDBMS are here to stay, as mainframes did (they never disappeared as it has been predicted by many "experts" in the past). New technologies never replace good old ones, they just complement them.

Anyway, there are tangible impacts of SOA and the Cloud on the database market:

We will access more kind of data sources in the future, not only RDBMS and services, but also new kind of databases. Heterogeneity will continue to grow.
We will access more data sources in the future, most applications were using single databases, they will now access multiple data sources. We are switching from data access to data integration (I tend to prefer the term adaptive mediation of information). Integration has to be done at a business level, not at the SQL one or XML one.
Many advanced features of databases engines (security, fault tolerance, stored procedures...) will progressively move to an intermediate integration layer. Databases, including relational databases, will go back to simple and efficient storage technology.
Data integration will become more important than the database itself, databases will be commoditized. Each application development team will be able to select the best database technology for its needs.
Accessing non-database data sources will impose to have extended metadata. The relational world is simple because SQL provides a convenient, technical APIs to access data at the atomic level (a cell at the cross between a row and a column). Everything is implicit, in terms of metadata, access patterns, etc. Conversely, accessing a service-oriented data source imposes to explicitly describe its data model and its data manipulation semantic. Services can be either fined-grained or coarse-grained, you need to capture that. Data access has its contribution to the Semantic Web.
When thousands of data sources will be available as data services (like mainframe screens, APIs of packaged applications), we will need tools to automatically combine them at runtime. Manual, hard-coded or even visual composition of data services is a choice only when dealing with a few data services. Dynamic composition of data services (e.g. aggregation of fine-grained data services into other larger coarse-grained data services as required by ever changing business functionalities) is imposed by the really agile IT. Otherwise "agile" will turn into "fragile"!
Ad-hoc data mashups will require availability of the right data services at the right time. This can only be achieved by platforms being able to dynamically create and publish new data services as they become required.
Access to non-structured data will grow. At the same time, non-structured data is on the way to structure itself or at least to describe itself better, see the "Linked Data", "OpenCalais" and the "Web of Data" efforts for instance.
Accessing multiple data sources with different latencies will impose to deal with reactive data integration patterns. We will have to support asynchronous data access, and we will need tools for that, because asynchronous and parallel programming are not natural to most developers and architects.

As Martin Fowler concludes, Data services platforms are enabling the promises of SOA, by really favoring small business functionalities having their own storage, instead of sharing data in huge centralized databases.

Business Objects Data Services

noreply@blogger.com (Eric Samson) — Mon, 10 Nov 2008 22:08:00 +0000

I've recently read several white papers about the new Data Services offer from Business Objects.

Seems to me, it is basically a renaming of their former ETL and Data Quality products.

It is not fundamentally surprizing to see ETL vendors moving to Data Services, as EII vendors did before them.
Let’s say that globally an ETL is a tool to move data from DB1 to DB2, or more exactly extract data from DB1, transform data somewhere (huge debate here) and then load data into DB2.
Now, let’s suppose you replace the third step by “publishing data”, you then have an "ETP" or even a Data Services Platform, if you publish resulting views as Web Services.

Well that's probably still targetting read-only, non real-time data integration, but at least it demonstrates that Data Services are gaining momemtum on the market.

SOA social

noreply@blogger.com (Eric Samson) — Sat, 08 Nov 2008 15:55:00 +0000

I found the article mentioned in the previous post on this portal -> SOA Social.
You'll find interesting resources over there.

The case for coordinated EDM and SOA

noreply@blogger.com (Eric Samson) — Sat, 08 Nov 2008 15:53:00 +0000

Article by Keith Worfolk in SOA World about the benefits of coordinated strategies for Enterprise Data Management and SOA.

Needless to say that a Data Services Platform should be the beating heart of these coordinated strategies.

I cannot agree more with the first best practice described by the author:
"...When thinking about services, don't forget to consider the data.
Systematically designing a service model is like designing a data model. For either, its impact should be considered long term, and the level of normalization of designed components, services, or data is considered a sign of quality and maturity.

Figure 6 shows service-data normalization from immature to mature organizations:

"Wild West": Non-existent or ad hoc and uncoordinated normalization
Ownership/Stewardship: Service designs built on data designs
Encapsulation: Service and data designs coordinated in development/maintenance initiatives; either may drive the other as long as they are coordinated
Object: One and the same service/data designs. Normalized designs are within EIA designs; service implementations take data ownership to another level where master data value is known only in service designs/implementations.

Most organizations pursuing services-data normalization have progressed to ownership/stewardship levels, yet need to reach encapsulation before realizing major benefits in efficiencies, maintenance costs, and asset business value.

The highest level of service-data normalization, object, may not make sense for some organizations, especially where master data or business services change frequently. Depending on their stability, the more possible an object level may be. However, cost/benefit analysis may make encapsulation preferred for some organizations.

Transitioning to advanced service-data normalization is a process of increasing organizational maturity toward coordinated EDM-SOA strategies..."

More on LINQ to SQL

noreply@blogger.com (Eric Samson) — Sat, 08 Nov 2008 15:33:00 +0000

Some additional comments about the possible end of LINQ to SQL in Julia Lerman's blog.

Data Mashups: Enabling Ad-Hoc Composite, Headless, Information Services

noreply@blogger.com (Eric Samson) — Fri, 07 Nov 2008 16:14:00 +0000

ZapThink just released a research paper about Data Mashups.
Extending the notion of mashups, data mashups will decouple data integration from heavy development cycles. But as ZapThink's Ron Schmelzer wrote this requires to have a strong Data Services Layer in place.

I fully agree with the following statements from the research:
"...the IT organization must give Service consumers the tools and methods they need to be able to successfully compose those Services with low cost and risk..."

And this exactly why Dynamic Data Services are so important. Having statically defined, hard-coded (or visually composed) data services could meet the requirements of statically defined service-oriented processes, but the reactive enterprise needs more flexibility. It is important to have the relevant data services available in real-time when data mashups will require ad-hoc data. A good Data Services Platform must support this kind of runtime generation and deployment of ad-hoc data services.

Later the author writes:
"...One of the important benefits of a Data Services layer is that it enables loose coupling between the applications using the Data Services and the underlying data source providers. Loose coupling enables data architects to modify, combine, relocate, or even remove underlying data sources from the Data Services layer without requiring changes to the interfaces that the Data Services expose. As a result, IT can retain control over the structure of data while providing relevant information to the applications that need it. Over time, this increased flexibility eases the maintenance of enterprise applications..."

In a world where most data sources will become service-oriented (even the databases themselves), it is important to be able to really achieve the decoupling between the data services and the data sources. In this particular case, this requires extended semantic metadata around data services, so that an advanced Data Services Platforms can dynamically recompose them at runtime, as requested by new data mashups.