Operational BI (or Operational Reporting, the terms haven't really hardened yet) is a really important topic. The dialogue for the past decade and a half has been about data warehouses, loading data warehouses and using BI tools to report from data warehouses. It's become apparent that this is not a complete solution. Data warehouses typically (though not always) are designed to support after-the-fact reporting and analysis and the data models are oriented for that. Operational reporting is more immediate, more low level and more actionable.

There was an excellent webinar on BeyeNetwork today, sponsored by GoldenGate Software and delivered by Claudia Imhoff on this topic (an archived version should be available there by the time this blog is posted). Claudia explained the various options that exist (and the drawbacks of each), such as going directly against operational systems, using ODSs, re-architecting data warehouses to provide fresher data and more pliant data models and, presumably the most attractive solution, using various flavors and types of EII (Enterprise Information Integration) to dip into the data warehouse and the operational systems by creating some abstraction to allow data to appear as if were in a single location.

Why is Operational Reporting important? One reason is that the rise of business process modeling and execution systems will inevitably make business processes more dynamic and the data warehouse/BI structure will not be able to keep up. Pulling data directly from operational databases (or replicates or logs or even on-the-fly from message queues) avoids the steps of transformation and loading to a data warehouse, eliminating a lot of latency. In addition, data warehouse models are not operationally oriented, so its conceivable that in the transformation process, a lot of the semantics of the data at the operational level are lost. As a result, data created from Operational BI could not easily be written back to the operational systems.

Another reason is that BI has not been able to cross that "actionable" gap because its architecture is geared toward informing people, not operational systems. An Operational "reporting" system has to do more than just report, it has to react. That reaction can be driven by a human being (picking up the phone and sending the store manager to investigate a problem) or it can be driven by decision services as James Taylor and I described in Smart (Enough) Systems: How to Deliver Competitive Advantage by Automating Hidden Decisions.

Now that it is possible for business people to interactively modify running business processes with process automation tools, they will need help assessing what they've done. There won't be time to study the problem, devise some data models to support investigations and build data integration aplets. EII allows for a sort "canonical" map between the systems and analytical/reporting model so that these constant changes can be handled at a higher level of abstraction, almost instantaneously. By aligning the substantial resources of the BI industry to this effort, I'm pretty confident we'll make a lot progress quickly.

You can't fix a broken company by measuring how broken it is. You can't even improve a functioning company by measuring where it could do better.

The BI industry has sort of casually sent the message that BI makes companies better. I've seen this in presentations, webinars, seminars, books and blogs from vendors, practitioners and analysts. But the question is, once you expose something that needs attention, what next? As a consultant and implementer of data warehousing and BI for many years, I never really came up with a good answer. I always referred to it as the Jordan River problem: we could take them through the desert and to the east bank of the Jordan, but like Moses, we couldn't get them across. Even when we did everything we set out to do, when approaching management about the next phase of the operation, to help the client start addressing the problems with employee morale, high turnover, inventory snafus, poor customer service, etc., the response was usually something like, "Neil, aren't you the data warehouse guy? Shouldn't we get McKinsey in here to work on that?"

In short, if you're involved with BI, you may have as good or even better insight into what is going on in the company, but you clearly lack the portfolio to do anything about it.

Part of the problem is that we feel successful if we can just get people to USE what we've built, so maybe it's time to stretch a little. In our proposals to clients, we should measure our success by how effectively people use what we built, and develop the programs and consulting expertise to see to it that they do. The first step in doing that may be to address the issue of decision-making in addition to just informing decisions. I'll be generating some original research this year on decision-making, but it's only the beginning. If we can deliver BI with a better understanding of how decisions are made (or avoided), how good/bad decisions are made, we can more or less stay in our patch but offer a much better service.

This practice could include things like a decision audit, workshops for helping people understand how they really are deciding things (versus their mental model which may be incorrect) and, finally, tracking actual decisions and evaluating how well they were made. Now THAT would be Business Intelligence. All of this can happen if we can ride BI into the areas of the company that need attention. If we perpetuate the BI-is-an-IT-project mentality, or that BI is for go-to-guys, we won't have the wormhole that moves us past the McKinsey barrier. But if we do, those decision makers that can change the way the company works will be our clients. 

Anyone want to join me in building this competency?

Bill Inmon posted an article  where he discussed the drawbacks of a "virtual data warehosue." Now the idea of a virtual data warehouse has been around for years (I remember IBI positioning their EDASQL technology as a virtual data warehouse fifteen years ago). That was a terrible idea then because it provided only a method to access data, with no proper cleansing, integration or persistent storage of it. It was, frankly, not a data warehouse at all. Using that definition, anyone would agree that a "virtual data warehouse" was not a workable solution.

The problem is that Inmon defines any method to augment the data warehouse with access to source data directly as a virtual data warehouse, and in the same terms as the IBI concept from long ago, citing its shortcomings, casting the virtual data warehouse as not only a bad idea, but one that has been historically bad and discredited. In fact, the term "virtual data warehouse" has been used for years by detractors for any idea that augmented or supplemented the data warehouse architecture, many of which were truly misguided ideas. In today's world, though, it is not only possible, it is necessary to supplement the data warehouse. This "virtual data warehouse" is a new thing, something not possible before.

Inmon's primary argument about virtual data warehousing is that it skips the crucial step of integrating data:

The great appeal of the virtual data warehouse is that you do not have to face the problem of integration. 

This is the central premise of his argument, that a virtual data warehouse skips data integration.  What actually happens in today's version of a virtual data warehouse (and it shouldn't really be called that) is that a central data warehouse endures, with the kind of integrated data that Inmon would approve of. However, since there are so many data sources now, they change more frequently and the volumes are extreme, the batch load data warehouse is burdened with latency and often impractical for 100% of the analytical requirements. Because system architecture has changed so radically over 15 years, it is possible to read logs, queues, perform changed data capture and even go directly into the operational systems without resorting to extraction into staging areas or degrading their performance. The semantics of these source systems are considerably more transparent than they used to be and the speed of processing is so much faster it is possible to cleanse and integrate data on the fly. In those cases where it is not feasible, or where for a myriad of reasons it makes sense to warehouse the data, that process is preserved. So the virtual data warehouse is actually more of a surround strategy than an alternative. 

Why then is the virtual data warehouse such a supremely bad idea? There are actually lots of reasons for the vacuity of virtue manifested by the virtual data warehouse. 

I won't repeat all of his reasons, but they all seem to be related to the old "it will bring the system to its knees" claims about the inefficiency of federated queries and the affect on systems and network performance. I call this reasoning "managing from scarcity," except there is no scarcity anymore. Again, not every query is federated; many will be satisfied by the data warehouse. We can build systems thousands of times larger and millions of times faster than we could two decades ago as data warehousing was catching on and isolation of data and processing was absolutely necessary. Today's applications and business processes call for faster, fresher data than a data warehouse can provide, in many cases. Rather than ignore these requirements, a virtual data warehouse (with adequate semantic rationalization) gives the developers a chance to position the processing to the logical location, and the ability to use abstraction (or indirection if you prefer that word) to address the data, providing the flexibility to change that configuration whenever necessary. 

The ugly truth is that the corporate analyst MUST do integration whether he or she wants to or not. There is no getting around it. Wishful thinking just does not cut it here.

Who wouldn't agree with that? The question is, where and when? The bulk of the difficult integration work is manual - people gathering information and getting buy-in that they have the right mapping. Where that process gets implemented is not the issue. There is no rule that says it must all go to a data warehouse. Historically, that was the practice because of the constraint of resource limitations that are now mostly relaxed.

So the problems with the virtual data warehouse are legion. There is poor performance or no performance at all. There is an enormous system overhead. There is only a limited amount of historical data. There is the work of integration that each analyst must do, one way or the other. There is no reconcilability of data. There is no single version of the truth for the corporation.

All of these things are true in the virtual data warehouse the way Inmon defines it, a big mess of federated queries directly to source systems with no integration, no history, no real context. No one would suggest this is a good idea. But a data warehouse today should be a participant in a constellation of source data, so long as the access methods to data beyond the warehouse's control can be relied upon to be correct, timely and not a burden on other systems. This kind of arrangement is not a virtual data warehouse, it's an analytical framework. We should discuss the merits of the idea as such, not dismiss them because we name them with a long-discredited concept.