As an architect...

I have been playing the role of a DW Architect for a large financial services organization over the past 6 months. When I reflect back on my duties and responsibilities, one aspect came out very strong that gets missed out in the job description.

"Leaves his Unique Selling Proposition trace in the project"

There are often a lot of other crap that the organizations look for like..Coordination between the business and IT....Owner of data models....Provides strategic direction to the IT team...etc...

Do you think Emperor Shah Jahan would have given a big job description to his architect, Ustad Ahmad Lahauri, when he set out to build the Taj Mahal? He described the Taj Mahal as

Should guilty seek asylum here,
Like one pardoned, he becomes free from sin.
Should a sinner make his way to this mansion,
All his past sins are to be washed away.
The sight of this mansion creates sorrowing sighs;
And the sun and the moon shed tears from their eyes.
In this world this edifice has been made;
To display thereby the creator's glory.

If you read the last line, it stresses the need to display the creator's glory. An architect should leave behind his glory after he is long gone from the project. That should be the true job description. It should be the one thing that the people would still talk about after he has resigned. The rest of the responsibilities are just enablers for the ultimate glory.

Personal Analytics - Excel & Qlikview

Everyday, we take decisions, either at work or at home or during transit. How do we take our decisions? When we take decisions at work, we use "saved" data to help us out. But at home, we rely on our primary memory to help us out. We also rely on reviews on the web to help us take our decision. We rely on "Facebook" likes. But somehow, I am still not convinced that this will help us to make a decision that will be the closest to the best. So I decided to maintain my own personal analytics database. My first "personal" problem was finding the right health insurance for my dad. So I decided to make my own BI environment to help me take this decision.

1. I used Excel for feeding & maintaining data. I profiled attributes of all kinds for this.
2. I used Qlikview for BI analytics.

It is turning out to be a great exercise for me to find which insurance is the right one for my dad.

Metadata Management - Scratching your own itch

Meta-data management is always a complex problem, because its about capturing "data about data". I don't think DW as an industry leader is anywhere near on making sure organizations are cleansed of bad data. We still have bad data and if it is about capturing finer details about this "bad data", its even more bad.

My current customer has this huge problem of not knowing how the data elements get mapped between different hops in the data warehouse (Staging, DW, Datamarts, Business Objects Universe and ultimately the "Requirements"). There were lots of discussions carried on what kind of tools to procure and the profiles of meta-data architect to recruit. We were getting nowhere.

We decided to scratch our own itch. We started this exercise 3 months back in our past time to start documenting the data lineage in a simple "denormalized" spreadsheet. It took us time. Layer by layer, source by source, we did it. After 3 months, we had a full blown spreadsheet which captured the complete data lineage and the business rules implemented in the DW layers.

When we reflected back, we realized a couple of "eye-openers"

1. Don't invest upfront in meta-data. Ask your existing team to start documenting in the easiest and most flexible manner possible. Probably a spreadsheet. Take it one step at a time.

2. And start small. Get the data first and then think about tools. Check if the data helps you to make your job any easier or the business user's job any productive. If not, the meta-data program is not for you.

Design Documentation - Batons in a relay race

Documentation is always like a "baton" in a relay race. You care for it, only when you pass it over to somebody else, till that time you don't even bother of the existence of such an entity. If you have to make a baton interesting to a running athlete, you need to stuff it with something that is interesting to him; probably some sort of energy drink, which he will consume it whenever he becomes exhausted. The beauty of the baton design is that it has to be light enough so that its not an extra burden to the athlete and designed in such a manner so that it enables for a quicker passover between the athletes.

I had always stayed away from "Documentation", because I never found a usage for it except during audit or knowledge transition. And even in audit, nobody cares for the quality of the document; the auditors just check for the existence of the document.

So to make some sense of this complex phase in an SDLC, I decided to derive the "Baton" analogy to documentation, because just like a baton, without any documentation, I can never say that somebody completed a relay race. To make documentation competitive, I decided to do the following: Make it -

1. Light - so that its light and easy to carry.
2. Interesting - so that the consumer opens it and uses it frequently (and)
3. Do its job - so that the transition/passover is easy

Light

Why is iPad 2 thinner and lighter than iPad? Simple. They moved from 2 thicker batteries to 3 thinner batteries. They made the whole shell using carbon fiber. Not that I understand the material composition of carbon fibers to comment on it, but if you reduce the content, it becomes lighter. So essentially, I will try to make my document as light as possible. "Fewer pages" will be my KPI. So I decided to budget a page count for every kind of document and focus to convey whatever I wanted to convey within that page count. And I have a motherhood rule "Don't cross 15 pages of content".

Interesting

For anything to be interesting, it should be useful. Information in the document should be useful and should in turn help the consumer/creator of the document. So to make it interesting, I decided to add 3 simple sub-sections for every section that I created - What, Why & How ? So, if I have to design Change Data Capture in my ETL (some data warehousing terminologies) process, then I add What is CDC; Why should I use CDC?; And how should I enable CDC?

Do its job

It should do its job. It should enable transition. So, if the successor wants to understand what the predecessor did, the document should be able to convey. It shouldn't be lost.

So if a design document addresses the above 3 philosophies, it has met its purpose of existence.

Can BI Strategy ever become real?

When a super BI consultant recommends a whole stack of To-Do items in the information curve, how real is his recommendation? How real is the BI Strategy? 30%, 50% or 70%. Have you as a customer asked your vendor about it? Ask it. The typical response would be - "Depends".

The problem is not with the BI consultant; the problem is with prediction. BI is such a game where the variables are too many - money, business benefit, pain problems, information maturity, tool consolidation, vendor proliferation, data volumes, system integrators, application support, advanced visualization, data conformance, data quality, stewardship, etc....The list just goes on and on. BI Strategy almost turns into a weather forecasting system. So is there no answer to being real? There is. Answer is "Stop doing it. Get Real."

BI comes with a cost. Its not something that you can purchase it during a sale. BI is something that every organization needs. It has become ubiquitous. A strategy is just a sales tool to your governance board for approval. Do you need one? Why do you want to spend on a sales tool to prove that it is required for your organization? Would you construct a business case for seeking an admission for your son or daughter into the IIMs or the MITs of the world. Instead spend it on building a 60-day data mart. Make the users use it for a month. After a month, pull the plug off. The # of calls you recieve to get the system back would talk about the ROI of BI.

Let me know what you think.

Operational BI - Part 2

Having set the need for an Operational BI in my previous post, I will sketch out the architecture of an Operational BI solution.

The four important blocks to be considered while designing an O-BI system are

1. Sourcing/Extraction Module
2. Transformation & Load Module
3. Data Retention Module
4. Reporting Module

Sourcing/Extraction Module discusses the extraction strategy from the source. This Module covers the change data capture (CDC) design & data transfer mechanism. Choosing the right extraction strategy would spell success or failure for the project. Almost 70% of O-BI projects fail because of the wrong sourcing strategy implemented.

Transformation & Load Module discusses the kind of loading tool-set that would suit an O-BI system. Details about the expected load volumes, the loading patterns and the hand-shaking mechanisms with the source will be discussed

Data Retention Module discusses about the parameters required for estimating the size of the sliding data storage windows.

And finally the Reporting Module discusses the kind of reports that an operational executive would need for taking his tactical decision on a hour-hour basis.

These sections would be discussed in detailed in my further posts.

Operational BI - Part 1

The genesis of BI has always been the need to seek for the BIBLE of decision making. But BI over a decade has transformed itself from a night watchman to more of a 24/7 call-center representative. It has become real-time. What made this change? Why was the mutation to real-time necessary? What are the challenges in data integration? And finally, how can Operational BI (O-BI) be coupled with the Enterprise Analytical Reporting framework? I will be assessing each of the questions posed in greater detail and arrive at a design pattern for modeling a Operational BI Solution.

Let us drive the need for implementation of an operational BI solution with an example.

A store manager at a retail outlet manages various aspects of retailing - visual merchandising, customer experience, resource scheduling, loss prevention, product management (ordering, receiving, pricing, inventory). Let me explain each one of these facets of the retailing business briefly.

• Visual Merchandising: Promotion of the sale of good through visual appeal in the stores (source: Wikipedia).

• Customer Experience: Reduced customer wait-time in the check-out counters.

• Resource Scheduling: Monitoring the efficiency of the employee schedule for improved load balance of employee work-hours.

• Loss prevention: Real-time monitoring of 'shrinkage' because of shoplifting, employee embezzlement, credit card fraud, system errors and many more.

• Product Management: Real-time monitoring of product inventory.

Given this background, I would proceed on to connect all these process areas with a business case that would put a Store Manager in trouble and how an Operational BI solution can save his day.

Let us assume that the Store Manager has access to a reporting solution which refreshes once in a day. He notices that the daily sales has dropped as compared to the previous day. He drills further down to investigate the cause of the decline. He finds out that the drop can be traced to one particular hour in the day. A deeper look into the problem highlighted the issue of an increased average customer wait-time per hour causing a poor conversion rate. The wait time finally was attributed to reduced work-force in that hour because of an increased lunch break taken by the employees (since they turned up very early to work).

This problem could have been easily rectified if the store manager had access to data earlier than what he had. Had he had real-time access, he would have noticed the dip in sales for that hour immediately and would have taken corrective action, thereby not affecting the sales during that hour. With a decent business case established for a real-time BI system, let's analyse what an operational BI is and how does it facilitate to solve the problem.

The architecture of Operational BI and the challenges associated with it will be posted in the next article.

Infobright's column based datawarehousing

Infobright, a open source data warehousing start up addresses the performance issues that usually come along with a data warehouse by implementing a highly compressed column-oriented store. The data is stored in columns instead of rows. This allows for reduced I/O because of the compression ratios obtained on the columns. Data is stored as 65K blocks or nodes containing a meta data store about the relationships between columns.

Some of the key customers of Infobright are RBC Royal Bank and Xerox. They claim their product would be ideal for data warehouses ranging from 500GB to 30TB. Their compression ratios are close to 40:1 according to their community blogs. The most attractive feature about them was the compatibility with the existing Business Intelligence tools like Business Objects and Pentaho.

I wasn't very convinced with the concurrency offered by them. It supports 50-100 users with 5-10 concurrent queries. I will watch for the progress of this new exciting player in the already crowded BI market.

Release Early Release Often

Release Early Release Often (RERO) technique proposes to have releases early and often, instead of a big bang release. This approach is typically followed in tech startups, working on Open source projects. That’s the reason we see many of Google’s products still in beta version and their updates getting released once in a month or so. We planned to experiment the strategy for a big Master Data Management (MDM) project. The experimentation turned out to be successful. The rest of the essay discusses the experience details of such an implementation.

User Thrill

Important features of the application were phased out for various distinct releases. Some of them were Hierarchy & Workflow management, Security and Exception reporting. And the duration between releases were as close as 2 weeks. That meant, the user saw features getting added once in 2 weeks. We captured the user feedback about the releases and made sure we corrected it in the immediate ones. This approach had a two prong benefit. User experienced the application very, very early and we experienced the bugs. By the time, the UAT phase reached us, the application had reached a near-to-zero defect zone. We were a bit skeptical whether the user participation would be high, but since the product was there to be played with, it naturally attracted them.

Incremental Application testing

The application was getting tested from the day the first beta was released; rather from the “Go Live” day. Although this created few negative impressions on the user experience due to few unpleasant bugs; they knew that it was in its beta stages and the next release would have the patched version. In fact, our testing team grew from a 3 member team to a 6 member virtual team (There were 3 business users).

Support framework

To enable such a dynamic release process, the revision control and the code review/release systems should be efficient; there would be multiple releases instead of one. The integration testing should be solid. And the unit testing before the releases should be good enough not to distract your users completely; dissolving the purpose. Meticulous planning of the releases will also form a key to the success. The development tools that you use should be agile and adaptable enough to accept and implement the user’s feedback for the next release.

Conclusion

The experiment turned out to be a success. This strategy would work for most of your implementations, unless it’s a maintenance project with less than a week’s duration of deliverable.

Which MDM approach is right for you?

MDM, in the past 5 years, has come a long way in its maturity model. Most of the MDM implementations fall under 2 different kinds of approaches.

1. Operational MDM (the tougher among the two)
2. Analytical MDM

Operational MDM enables synchronization of master entities and their attributes between the transaction processing systems. Why does one need such an MDM? Let's take an example. ABC Corporation is a manufacturing firm. It conducts roadshows and marketing campaigns to advertise its products. The salesperson collect customer information during those roadshows and feed it into their IT systems for further followup. There are a different set of sales representatives who conduct feedback on their products sold, with their customers. They too enter the customer feedback into their IT systems. These are 2 different sets of CRM processes.

Typically what happens in a mature company is, there are a set of batch processes which pick up the master data from one system and transfer it to the other. Now this introduces delay, inconsistency, inaccuracy of data and lot of manual reconciliation (same customer name can be entered by 2 different salesperson or the latest survey from a salesperson can erase previously collected information about the customer). So the IT develops custom programs to clean up the data, write reconciliation programs but still cannot manage to do all this in real time.

This mess can be reduced or eliminated by deploying an operational MDM. Operational MDM tools solve the synchronization problem using complex match-merge algorthims. Some of the tools currently in the market are Siperian, IBM, Purisma, Oracle and SAP.

Analytical MDM is an architectural approach if the problem revolves around inconsistent reporting for business performance management. In simple terms, inconsistent hierarchies are getting reported out. This needs for a unified reporting view of the master data. The audience for this system would be the downstream data warehousing and business intelligence applications. Some of the MDM vendors selling their expertise in this area are Kalido, Oracle, IBM.

It is essential that an organization has to build both these models to address their MDM needs. But which one to chose first depends on which problem is in their high priority list.

Teradata's reseller alliance with Trillium

Teradata Corporation announced its reseller alliance partnership with Trillium Software. Teradata will now combine its warehouse product with Trillium's Data quality tools and its own MDM products. Overall, this seems to be a good strategy for Teradata, because now Teradata's customers can leverage Trillium's data quality abilities on their huge databases.

Because of this alliance, the customers will enjoy a powerpacked database, Data Quality tools and a MDM suite. Information Difference has ranked Teradata's MDM low in the quadrant though compared to the likes of SAP, Oracle and Siperian.

Buy or Make - Financial Analytics

Today, I had a consulting assignment with a company focussing on Server Virtualization. The objective was to narrate the factors influencing a Make vs Buy (mVb) decision and their risk quotients for a Financial Analytics Solution.

Some of them are :

1. What is the business requirement and is the requirement very unique?
2. How urgent is the application?
3. What is the technology Strategy of the Organization?
4. Does the off-the-shelf product address most of the requirements and does it have flexibility to customize it?
5. How does the present make-buy decision relate to the strategy?
6. Are their right people and support systems to support the application, in case of a build?
7. Does the financial tool address internationalization needs?
8. Are their security measures in-built in the tool, because it hosts sensitive data?
9. Can the Integration of the Packaged Solution into the process control system be done seamlessly?
10. What is the underlying technology? In this case, what is the ERP system? It would make sense to buy the analytical solution from the same vendor of the ERP system, if it addresses your requirements
11. Will the TCO be reduced because of the Buy approach?
12. Are their right people and support systems to support the application, in case of a build?
13. Does it reduce cost?

After these questions were answered, the following matrices were prepared which summed up the decision.

1. High Level Requirement x Priority x Effort Estimation Matrix
2. Benefit Comparison Matrix
3. Risk Comparison Matrix

Statistics and Data

I was reading an excellent text "Statistics for Business and Economics" written by Anderson & Sweeney. It highlights the importance of statistical measures in decision making. Many of the existing predictive analytical tools use most of the principles covered in the text. It also highlights the importance of collecting and preserving data.

One such example covered was to calculate the average wait time of a queue in a particular ATM in New York. Using this data, the bank would then decide to position a new ATM to balance the load in that busy place. The predictive model uses probability distribution and helps the analyst in making a decision. The models have to be refined so that they don't reflect any false positives.

Why should I make my MDM SoA enabled?

A Car Sales Manager is capturing the details of a customer, who visited his showroom. After jotting down the client's address details, the Manager wants to check out if the address is a valid one. How can he achieve it?

A PRO in the same organization receives a call from the customer that he wants to change his address in the system records. The PRO logs in to the silo-ed application and enters the new address. When entering the address, the PRO wants to check if the new address is a valid one. How can she achieve it?

MDM and SoA make this happen. MDM is more like a service provider and SoA is a framework helping the consumers to access the service with ease. A Location Master Repository tied up a SoA framework makes any consumer to use the services of the Location Master. This way, the same master data gets reused throughout the organization for multiple purposes.

In one of our projects, the client had a unique requirement to assess if a product is worth promoting and if found promotable, what is the promotion to be given to it. Such a requirement was addressed using a SoA framework on a product portfolio MDM. The architecture had BEA AquaLogic Service Bus interacting with Kalido MDM to provide the services. The challenges lied in identifying the streams which would consume this service and evaluating if its worth considering a service. The service found its usage in many CRM applications in the organization.

MDM - SoA Marriage

After a long break from blogging, I am starting my series of explorations with MDM and SoA. When would somebody go for a SoA implementation for an MDM solution? Is it because it's a enterprise-wide initiative to make everything the SoA way? My current project proved to be a big failure on this front. We had to build an MDM solution and the enterprise architecture team had a clear focus to make anything and everything SoA enabled. The MDM solution was built on Kalido. Who are the consumers of this MDM data? Answer is . A series of downstream CRM applications. Sounds good. Where would the SoA architecture fit in? Is it just in the Consumer world or also in the Sourcing world? We had to design the reporting solution completely SoA. But during a period of stress testing, it proved that the SoA framework just couldn't handle the volumes the downstream applications were streaming. The MDM Solution would get a huge number of updates from the ERP stack everyday and all these changes had to be donated to the consumers. The SOAP message was just too big to be parsed by the reporting solution. The users had to wait for a considerable amount of time to get their reports out.

So my question is whether you design the MDM solution the SoA way expecting in future that things with performance would get solved or wait till things get solved and then re-architect the solution?

Currently, we have the SoA suite disabled and reports are being fired out from the databases directly to the reporting solution.

XML-based MDM

After a brief hiatus, I am writing this article on Orchestra networks' EBX.Platform - an XML based approach to Master Data Management. The EBX.Platform is based on J2EE and XML. The whole architecture is based on 3 items - Models, Services and Modules (Models + Services).

So how are they able to achieve their MDM framework?

The Data Model is developed based on a simple XML schema standard and this is termed as an Adaptation Model. Services such as import/export can be added on top of the adaptation models. It can also be maintenance features provided by the UI. And finally Modules, nothing but Models and Services are deployed as Web applications.

They also support quite a unique feature - Branches & versions of Master Data. This helps a company to maintain its current version of master data, when it is working on a futuristic model.

I would like to monitor the progress of this interesting tool, given that it is being used in some big companies.

Data Modeling a Maze

Couple of weeks back, my friend took me to a maze. I was lost in a couple of minutes and was getting really frustrated after a while. I wasn't sure what algorithm they had used to construct the maze.

The only algorithm that I knew was the "Wall follower". All you have to do is it to follow either your right-hand or left hand touching the wall and you will reach either the exit or the entrance. I did take the longest path, but eventually reached the EXIT.

This algorithm would work only if all the walls are connected to form a loop. From that point, I was quite fascinated with the algorithms associated with the maze. There are also a few other efficient algorithms like Tremaux's algorithm. Visit Think Labyrinth for more fun.

After that, I thought of simulating a maze. Unfortunately, I am quite inept at programming languages, so decided to do what I know best. I thought of creating a simple data model for a Wall Follower Maze. It turned out to be quite an interesting problem. It took me around 30 minutes to come up with a decent logical data model, that would work for quite some scenarios.

So, the first model that I came up with is shown below. (Click on the picture to enlarge)

Let me give a quick explanation of the model

1. Design Co-ordinate: Super-type Entity
2. Entry, Exit and In-Maze Coordinate: Sub-Type Entities of DESIGN CO-ORDINATE which contains the co-ordinates of the location, where a player has to take a decision.
3. Decision: Entity which holds whether to turn LEFT,RIGHT,UP,DOWN or ABORT.
4. Decision Map: Entity which holds the map of a START-COORDINATE, DECISION TAKEN(whether to move left,right,up,down (or) abort) and an END-COORDINATE (the co-ordinate where he lands after he takes the decision).
5. Player: Entity which holds information about the player of the maze.
6. Movement: Entity which tracks the movement of the player.

There is one interesting phenomenon happening in this model. If an intelligent player has to play this model, this model would work, because the association between MOVEMENT-PLAYER-DECISION MAP has been modeled as an Identifying relationship. It means if a player, tries to navigate the same path twice, the system will spit out an error, (simulating an INTELLIGENT player, because he would never do the same mistake again).

But if a DUMB player had to play this maze, then this model wouldn't work, because a DUMB player would make the mistake of traversing a path with no fruits, again and again. So the association should be made as a Non-Identifying relationship.

How can I model both the scenarios at one-shot, without introducing redundancy in the entities or associations?

One of the ways to model both the scenarios is to track both their movements as 2 different MOVEMENT entities and include the constraint in the INTELLIGENT PLAYER MOVEMENT's entity. But this introduces an extra associative entity. It's easy for a toggle situation, Yes or No, DUMB or INTELLIGENT.

But, if I were to model differing levels of intelligence, how would I do it in a data model, without writing any procedural code to do it? How can E-R data models be efficiently designed for Fuzzy Logic Systems?

I found this as an interesting exercise to showcase that E-R data models are still long way from being truly a self sufficient tool.

We need a modern day E.F.Codd

Kalido's Business Information Modeler

Today, I received an update from Kalido on the Business Information Modeler Engine. This is what Kalido claims about the product.

"Kalido Business Information Modeler provides a graphical design interface that can be used to develop and refine business requirements for new and existing information. Instead of modeling data and their structures, the Kalido Business Information Modeler allows you to model the actual parts of your business; customers, products, assets, transactions, even people – and define how you want to see information in context. Even better, the Kalido Business Information Modeler can be used to change and update your model directly against the Kalido Dynamic Information Warehouse, allowing you ultimate flexibility in meeting the information needs of your business. The Kalido Business Information Modeler dramatically improves your ability to meet the needs of your business when it requires it – not when how it’s stored determines it".

The product is due for March 2008. I am waiting to experiment on the new features it claims. I will be evaluating the product on the following questions.

1. Can an in-house data warehouse be easily migrated into Kalido?
2. Will the business layer completely abstract the data layer?
3. Is it just a visual aid for creating/maintaining your data model?
4. Will the data in the warehouse be used by the tool to help the modeler provide real-time feedback on the errors and the inconsistencies of the new model that he plans to implement?

I will be writing more on this interesting product after I get a practical hands-on. Visit www.kalido.com for more details.

Oracle snaps up BEA systems

Oracle has recently purchased BEA systems for 8.5 billion USD. One of the reasons that I think could be behind the motive of this purchase is that they wanted to get hold of the large customer base that BEA had. This will also help Oracle to compete with IBM tightly in the middle-ware space.

This move also helps Oracle's customers to move into a subscriber based model. Oracle claims that the vision of the acquisition is expected to accelerate innovation by bringing together two companies with a common vision of a modern service-oriented architecture (SOA) infrastructure and will further increase the value that Oracle delivers to its customers and partners.

Oracle has eliminated a strong commercial rival and forayed itself into into the enterprise middle-ware market edging to be the market leader.

Styles of MDM framework

MDM can be essentially fitted into 3 styles of framework

1. Registry based approach: The MDM contains a reference to the actual data stores and doesn't contain the data itself. It has pointers to the respective source systems for the attributes it hosts. Data governance and integrity are left to the source systems to handle. Quick way to setup a MDM. The registry will decide where to pick up the data from at run-time. I haven't seen many companies implementing this model.

2. Centralized Hub: The master data is integrated from different applications, cleansed, standardized, corporate-governed, secured, authorized and published to different subscribers from one central repository. This hosts the entire MDM data. Takes a long time to setup, but its one of the most efficient systems of integrating master data.

3. De-Centralized Regional Hub: This is similar to the previous implementation, but the corporate data is maintained in a global MDM hub and the regional/ business MDM requirements are maintained in a local/regional hub. This clearly differentiates the corporate from the regional needs.

Choosing which model to vote for is one of the key elements for the success of an MDM project.