OSGi: The Value Proposition?

In a recent blog Hal Hildebrand argues OSGi's value proposition in terms of its ability to reduce long term 'complexity'. Hal argues that whilst it may be harder to start with OSGi as it initially appears more complex, for large applications and large teams it is ultimately simpler because the architecture is 'modular'. A diagram along the lines of the following is used to emphasis the point.

Complexity over time?


As an ex-Physicists, I'm naturally interested in concepts such as 'Complexity', 'Information' and 'Entropy'; and while I agree with Hal's sentiments, I feel uneasy when the 'complexity' word is used within such broad brush general arguments. Indeed; I find myself asking, in what way is a modular system 'simpler'? Surely a modular system exposes previously hidden internal structure, and while this is 'necessary complexity' (i.e information describing the dependencies in the composite system), the system is never-the-less visibly more complex!

For those interested, the following discussion between physicists at a Perimeter Institute seminar concerning 'information' is amusing, illuminating and demonstrates just how difficult such concepts can be.

Before attempting to phrase my response, I visited Kirk Knoernschild's blog - IMO one of the industries leading experts in modularisation - to see what he had to say on the subject.

Sure enough Kirk states the following:

"As we refactor a coarse-grained and heavyweight module to something finer-grained and lighter weight, we’re faced with a set of tradeoffs. In addition to increased reusability, our understanding of the system architecture increases! We have the ability to visualize subsytems and identify the impact of change at a higher level of abstraction beyond just classes. In the example, grouping all classes into a single module may isolate change to only a single module, but understanding the impact of change is more difficult. With modules, we not only can assess the impact of change among classes, but modules, as well."

Hence, Kirk would seem to agree. As one modularises an application, complexity increases in the form of exposed structural dependencies. Note that one must be careful not to confuse this 'necessary' complexity with accidental complexity; a subject of previous blog entries of mine - see Complexity Part I & Part II

OSGi - Preventing 'System Rot'?

Those who have worked in a large enterprise environment will know that systems tend to 'rot' over time. Contributing factors are many and varied but usually include:

• Structural knowledge is lost as key developers and architects leave the organisation.
• Documentation missing and / or inadequate.
• The inability to effectively re-factor the system in response to changing business requirements.

The third issue is really a 'derivative' of the others: As application structure is poorly understood, accidental complexity is introduced over time as non-optimal changes are made.

Hence, rather than trying to frame OSGi's value proposition arguments in terms of 'complexity' - OSGi's value is perhaps more apparent when framed in terms of 'necessary information' required to manage and change systems over time?

Structural information loss over time for modular and non-modular System.

Unlike a traditional system, the structure of a modular System is always defined: The structural information exposed by a correctly modularised system being the necessary information (necessary complexity) required for the long term maintenance of that System.

In principle, at each point in time:

• The components used within the System are known
• The dependencies between these components are known
• The impact of changing a component is understood

However, the value of this additional information is a function of the tooling available to the developer and the sophistication of the target runtime environment.

The Challenge: Simplifying while preserving Flexibility

Effectively dealing with both module and context dependencies is key to realizing OSGi's true value in the enterprise.

To quote Kirk yet again:

"Unfortunately, if modules become too lightweight and fine-grained we’re faced with the dilemma of an explosion in module and context dependencies. Modules depend on other modules and require extensive configuration to deal with context dependencies! Overall, as the number of dependencies increase, modules become more complex and difficult to use, leading us to the corollary we presented in Reuse: Is the Dream Dead:"

The issue of module dependency management is well understood. Development tooling initiatives are underway to ease module dependency management during the development process; an example of which being the SIGIL project recently donated by Paremus to the Apache Felix.

However, Kirk's comment with respect to 'context dependencies' remain mostly unheard.

From a run time perspective vendors and early adopters currently adopt one of the following two strategies:

• Explicit management of all components: Dependency resolution is 'frozen in' at development time. All required bundles, or a list of required bundles, are deployed to each runtime node in the target runtime environment; i.e. operations are fully exposed to the structural dependencies / complexities of the application

• Use of an opaque deployment artifact: Dependency resolution is again 'frozen in' at development time. Here the application is 'assembled' at development time and released as a static opaque blob into the production environment. Operations interact with this release artifact, much like today's legacy applications. While the dependencies are masked, as the unit of deployment is the whole application, this decreases flexibility, and if one considers the 'Re-use Release Equivalence Principle' partly negates OSGi's value proposition with respect to code re-use.

Both of these approaches fail with respect to Kirk's 'context dependencies'. As dependencies are 'frozen in' at development time there is no ability to manage 'context' dependencies at runtime. Should conditions in the runtime environment for whatever reason require a structural change; a complete manual re-release process must be triggered. With these approaches, operational day to day management will at best remain painful.

In contrast, leveraging our Nimble resolver technology Paremus pursue a different approach:

• The runtime environment - a 'Service Fabric' - is model driven. Operations release and interact with a running Service via its model representation; this an SCA description of the System. Amongst other advantages, this shields the operations staff from unnecessary structural information.
  
• The Service Fabric dynamically assembles each System resolving all modules AND context dependencies.
• Resolution policies may be used to control various aspects of the dynamic resolution process for each System; this providing a higher level policy based hook into runtime dependency management.
  
• The release artifacts are OSGi bundles and SCA System descriptions - conforming with the 're-use / release equivalence principle'.
• The inter-relationship between all OSGi bundles and all Systems with the Service Fabric may be easily deduced.

The result is a run time which is extremely flexible, promotes code-reuse, whilst is significantly easier to manage than traditional environments. OSGi is an important element, but the use of a high level structural description used in conjunction with the model driven runtime are also essential elements of this story.

OSGi: The Value Proposition?

The short answer really is - "it depends on how you use it"!

Without a doubt, many will naively engage with OSGi, and will unwittingly increase operational management complexity beyond any benefits achieved by application modularization; see 'OSGi here, there and everywhere'. However, for those that implement solutions that maximize flexibility and code-reuse, while minimizing management, OSGi's value proposition is substantial; and the runtime used is a critical factor in realising these benefits.

How Substantial?

To date my only benchmark is provided by an informal analysis made by a group of architects at a tier 1 Investment Bank in 2008. They estimated the potential OPEX cost saving per production application, assuming that it were replaced with a Service Fabric equivalent; for the purpose of this blog one may equate Service Fabric to adaptive, distributed OSGi runtime.

Cost savings in terms of

• application release efficiency.
  
• ongoing change management,
  
• fault diagnostics and resolution,
  
• efficiency savings through code re-use

were estimated. The final figure suggested a year on year OPEX saving of 60% per application. Somewhat surprised at the size of the estimate I've challenge the group on several occasions, each time the response was that the estimates were conservative.

To turn this into some real numbers - consider the following. A tier 1 investment bank may have as many as ~1000 applications; each application typically costing $1m per annum. Lets assume that only 30% of the applications are suitable for migrating to the new world - we're still looking at a year on year saving of $200m. Migration costs are not included in this, but these are short term expenses. Likewise neither are the cost savings realized by replacing legacy JEE Application Server and middleware with the Service Fabric solution.

As always - 'mileage may vary' - but never the less, quite a value proposition for OSGi!

How Nimble is your OSGi runtime?

Hands up all of you managing OSGi dependencies via an editable list of bundles. Easy isn't it! It just works right!?

Well actually - it 'just works' for a single application running in a small number of containers. From an enterprise perspective you are unintentionally contributing to an impending complexity meltdown; an explosion of dependency and configuration management issues. And if you are unfortunate enough to end up supporting your own composite creations, you may well end up envying the fate of Prometheus and rueing the day you learnt to code.

Possible harsh? But I'm not alone voicing this concern!

In his recent article "Reuse: Is the Dream Dead?", Kirk Knoernschild continues his efforts to educate the industry on the tensions between code 're-use' and 'simplicity of use'. Kirk argues that as you increase potential re-use via lightweight fine-grained components, the complexity of dependencies and necessary environmental configurations corresponding increase, so making these same components harder to use.

A simple concept, yet if unaddressed, an issue that will make your life as an enterprise developer increasing uncomfortable and help edge OSGi closer to that seemingly inevitable 'trough of disillusionment'.

Yet, from a development perspective the issue of dependency management is well understood.

Whilst initially found wanting, a number of projects now exist to address this; including the SIGIL eclipse plug-in which Paremus recently contributed to the Apache Felix project, (SIGIL leveraging Peter Krien's BND tool).

In contrast, the issue of dependency management in Production is less immediately obvious, its impact more profound and generally ignored.

* Will aspects of the runtime environment affect the runtime dependencies within the application?

* Will applications be isolated from each other, or might they run within the same JVM?

* How are the released artifacts subsequently managed in the production environment with respect to ongoing bundle dependency and version management?

Echoing Kirk's concerns, Robert Dunne started his presentation at OSGi DevCon Europe with the observation that; 'whilst modularity was good, its benefits are often undermined by dependency and configuration complexity'. The subject of Robert's presentation? The Paremus Nimble Resolver, which is our response to the concerns posed by Kirk.

Nimble is a high performance runtime dependency resolver. To deploy a composite application to a Nimble enabled runtime (i.e. the Paremus Service Fabric) one specifies:

* The root component of the artifact.

* And a set of associated policies and constraints.

Nimble then does the rest.

Presented with the 'root', Nimble dynamical constructs the target composite; ensuring that the structural dependencies are resolved in a manner consistent with both organizational policies and the runtime environment within which it finds itself.

Nimble's OSGi capabilities include:

* Fragment attachment policies.

* Optional import policies.

* Import version range narrowing.

* The ability to resolve dependencies on extender bundles (DS, 'classic' Spring, Spring DM, iPOJO).

With Nimble policies allowing:

* The configuration of selected extensions.

* Flexible constraint requirement -> capability matching.

* The ability to configure optional dependency resolution behaviors.

Not just OSGi, Nimble is a generic artifact resolver with a plug-able architecture. Any artifact type may be supported, with support currently available for:

* OSGi Bundles

* POJO's, 'classic' Spring & Spring DM

* WAR

* Configurations.

A Nimble enable runtime quite literally dynamically assembles all required runtime application and infrastructure service dependencies around the deployed business components! Specify a WAR artifact and Nimble will instantiate the appropriate Servlet engine dictated by runtime policy attached to the WAR; i.e. Tomcat or Jetty Sir? Specify a 'Configuration', and Nimble responds by installing the target of the configuration, and of-course its dependencies.

Nimble not only directly addresses Kirk's concerns, but goes on to radically transforms our understanding of the responsibilities and capabilities of next generation composite aware Service Platforms. But most importantly, Nimble was created to enable effect re-use whilst making life simpler for you and the organizations you work for.

Complexity - Part II: It all depends on the Question you ask!

I previously argued that the apparent complexity of a system varies dramatically with respect to the type of question you ask. The answer to one question may make a given system seem inordinately complex, yet ask another similar question, from a slightly different perspective, and the same system appears very simple.

Hence, it is the question that dictates where the line is drawn separating hidden and exposed system complexity.

Assume I want to deploy a set of services to an Enterprise. These services have specific runtime requirements and interdependencies. The usual question asked is...

• "What compute resources do I have, what are their individual configurations and capabilities?"
  

The response to which, an extensive list of resources and associated configurations/capabilities are presented, that now need analyzing. Like the positions of nodes in a lattice, the initial question, and subsequent answer, expose too much unnecessary information!

In contrast, if I ask,

• "Out of the cloud of potential resource which may or may not exist, what sub-set resources currently satisfies the following conditions?"
  

The response requires no further thought. Whilst I may never know the configuration of everything, I'll always know whether there are resources capable of servicing my stated requirements. As the response to the question is simple, and requires no effort on my part, I have no issue in re-asking the question as may times as required; this is essential, as the one thing I do know is that the environment WILL change!!

Re-visiting the lattice analogy.

Because it is simple to measure emergent macroscopic properties such a pressure, temperature and volume, it is easy to re-measure these and so deduce the relationship between them over time - e.g. Boyles Law. This would have been a significant challenge if the microscopic quantities of position, mass and velocity for each particle had been used instead!

Abstraction versus Virtualization?

Resource abstraction is different from resource virtualization. Whilst the latter attempts to represent a physical resource with a “virtual” equivalent, this equivalent emulating the attributes as the underlying entity, resource abstraction masks the complexity of the entity (physical or virtual), representing this resource via a simplified description. Resource abstraction and resource virtualization are orthogonal / complementary and interdependent.

To Conclude

• As systems become increasingly distributed and composed of an ever increasing number of moving parts - we need to step back from attempting a microscope description of the environment, and rather describe it in terms of its emergent macroscopic characteristics.
• We need to intelligently define the boundaries - the point at which microscopic behavior gives way to a more appropriate macroscopic view. Also don't be surprised if several boundaries exist.
  
• Dynamic service discovery / dynamic service provisioning / re-provisioning are fundamental - they are MUST HAVE core behaviors.
  
• So avoid all architectures and solutions that assume a static world, comprising of fixed immutable resources at known addresses; NB including wiring systems together via static immutable middleware services! Unfortunately the vast percentage of current software solutions, and the mindsets of the engineers that built them.
  

Build dynamic systems, manage them with respect to their macroscopic properties and the management / complexity issue vanishes. Conversely, if runtime complexity is a serious issue - it's about time you redesign / rebuilt your systems as no amount of traditional management software will save you.

Paremus colleagues continuing to flag articles to me, and partially restored vigor (New Year and all that) - have conspired to overcome my Q4-07 Blogger's Block.

I'll start by briefly para-phrasing the recent crop of virualization articles.

Virtualization is great because...

• You can increase data centre resource utilization.
• You can simply restart a service on a new physically platform, should the current physical platform fail.

However there may be some dark clouds on the horizon because...

• Resource / Dependency Management and Security are problematic
  
• Operational Risk may adversely affected
  

Well, I'll go to the foot of our stairs!

Isn't it obvious that, "visible" runtime complexity is increased by current forms of virtualization. Given this, it is surely no surprise that virtualization can negatively impact manageability, OPEX and, ironically, service availability?

Increasing server utilization at the expense of increased runtime complexity seems like a poor trade; especially if you remember that complexity is proportional to the number of skilled personnel required - and so OPEX. From a recent survey by Sun Microsystems ( Sun Survey ) it would appear that many CIO's would agree.

Meanwhile, we're told that:

"the IT industry will develop a new generation of management tools to address manageability and security issues created by virtualization. A great opportunity for start-ups and large IT companies alike".

Doubt this? Then check the current datacentre virtualization hype, and the number of VC funded companies in this market sector. Clearly the adage - "Identify the Pain - and sell them the Aspirin" is still in vogue with our VC friends. Unfortunately such strategies are at best simplistic; at worse, they demonstrate both the level of stupidity only achievable via a fully qualified MBA, and also the lemming behavior of the IT industry.

This time, the patient (Enterprise IT), really does need more than yet another, in a long sequence, of expensive Aspirins.

But perhaps the established IT vendors will address the problem?

Let's see. How many established vendors after 15 years of client server computing have enterprise management frameworks that are:

• Simple to Use
  
• Cost Effective
• Simple to deploy
• Address simple requirements like configuration management for software, server, storage and networks.

Its been a while since I've been involved in this area (i.e. HP Openview, Tivoli and the like), but I suspect the answer is still the same.

So what real hope is there for extending such solutions to address the new complications posed by the service virtualization?

I'll let the reader come to their own conclusions.