Numerous architectures, designs, and techniques are available to meet whatever your company's data requirements are.
When we talk about the storage of data and management of information, there is always the issue of where and how. Where do we store the data and how do we manage it? The question of using a distributed model or a centralized model has haunted IT professionals since the dawn of computers. Today, there are so many options and techniques available, it boggles the mind.
In reality, nearly all information systems and computing devices store data in a distributed fashion. A "hard disk drive" is actually made up of several platters, and each platter has some format mapped upon it. Personal computers on up to massive enterprise systems have multiple, if not hundreds of, disk units. Data is now stored on both solid-state "disk" as well as traditional magnetic "spinning" disk. Furthermore, relational database management systems separate and spread data across tables, rows, and columns. In other words, the act of data modeling, normalization, and deployment causes business entities to be distributed in some way. File systems store unstructured data across multiple directories, in multiple documents, and in multiple formats. Given all this, what are the concerns and the decision points around database architectures when building and implementing a business solution?
To start with, computer history and the associated trajectory of information technology play a major role in the discussion. In the early days, computing and storage techniques were obviously limited. A single disk drive only held so many 1s and 0s. When the device filled up, another one was required to continue. The same is true with addressing schemes. To find, access, and process data, the operating system has to keep track of exactly where the data is located—both on disk and in memory. If the storage subsystem or operating system is limited in the number of addresses it can use, there is a corresponding limitation in what can effectively be stored and identified. In essence, this is the importance of understanding the systems and OS architectures available today. For example, Windows systems are different from UNIX systems, and they are both fundamentally different from IBM i or z/OS systems. This is due to the underlying infrastructure, the address schemes, and the available storage management techniques. In a distributed scenario, if a single server has reached its limit, what happens when more computing power or space is needed? Either another server is required, or something must be pruned. Adding another server to the mix also requires overt planning, design, and management to determine what data and what work goes where. This "partitioning" of data and work distribution can be complex and human resource–intensive, especially if it is not performed by the machine. Now, to be sure, all computer systems have limitations. The point is, what are the limitations to growth, performance, and scalability? What are the techniques available to overcome those limits, as well as manage the possibly diverse and wide-ranging facilities?
Personal philosophy also plays an important role in the decision to centralize or distribute data. Some people are just more comfortable keeping all of their "stuff" in one pile. Others are satisfied with allowing things to be scattered hither and yon as long as the accounting works and nothing gets lost in the process. An illustration of this lies in the difference between the older generation of programmers and computer users and the newer generation we have today. Twenty or thirty years ago, centralized systems were all the rage. Everything was in one place, secured and shared. When personal computers arrived, by definition, information became more distributed. Now with more and more mobile devices being used, there is a new level of data distribution that has emerged. In effect, everyone is much more comfortable with the notion that data, even personal data, is all over the place.
Given this background, what is best for your organization today? The answer lies primarily in two places:
- Actual requirements
- Ability of a given system to meet those requirements
When wrestling with the choice, you must identify and understand both the business requirements and the technical requirements. You must also consider the limitations, growth path, and total cost of ownership of any potential solutions. Every system and every solution has limitations. The real question is, how do you overcome, or live with, them?
When it comes to database management systems, there are still two schools of thought: centralized or distributed. After all, Linux/UNIX /Windows platforms are still referred to as "distributed systems." It is clear today that systems are distributed not necessarily based on the current capabilities of the respective machine, but rather their historical use, as in PCs being distributed around the organization or company and then linked together in some way. Another philosophy borne out in options available today is the choice between "scale-up" versus "scale-out."
Scale up is based on the idea that as your data storage, data management, and data processing needs grow, the computer system (HW, OS, etc.) grow up with you. This tends to be relatively easy and non-disruptive, as represented by the specific architectures found in IBM System z running z/OS and IBM Power Systems running IBM i. Both of these systems were built from the ground up to be multi-user and upwardly scalable.
Starting with IBM's System/38, progressing through AS/400 and now IBM i, the ability to grow up has been paramount in the design and implementation of the business application system. Technology such as 48-bit (1978) and then 64-bit (1995) hardware, as well as a very large address space in conjunction with single-level storage, have proven to be a solid foundation to build a database upon. The unique Technology Independent Machine Interface (TIMI) allows new hardware and microcode technology to be slid underneath existing applications, databases, and programs with virtually no re-design and no re-development effort. The single-level storage model allows DB2 for i database objects to be created by the developer without regard for, or identification of, any physical location. In other words, the operating system hides and handles the data space and storage housekeeping tasks. All of these things help avoid the need to unnecessarily distribute data and data processing.
On the other hand, scale out is based on the idea that, as your data storage, data management, and data processing needs grow, the solution grows out by adding one or more new systems to the equation and then spreading the data and corresponding workload around. While the advent of blades, virtualization, and sophisticated management mechanisms can help minimize the effort, distributed systems tend to require a little more planning and administration. Linux/UNIX/Windows systems (or server farms) traditionally represent this realm. It is not uncommon to see multiple systems, servers, or nodes configured to handle a given database application instead of a single more robust system.
Again, the choice to grow up or grow out is likely based on history, system capabilities, and/or limitations as well as a personal philosophy or experience.
What about clouds? Where do they come in? Simply put, clouds and the accompanying virtualization technologies can allow the best of both worlds for things like big data and diverse information types. Multiple systems can be used to scale out for the ability to handle growth and varying workload, while at the same time appearing as if only one system is being used. For example, when using Google, Apple, or Amazon solutions, there is no view of the underlying components from the users or even the application developer's perspective (this is similar to IBM i, by the way). On the other hand, clever architects, engineers, and sophisticated software are employed to give this illusion. Someone or something must bring all the pieces together.
When designing and implementing a relational database, there is more to consider than just storage and compute nodes. Things like fit for purpose, data integrity, data availability, query performance, parallel processing, and latency are just some of the things that must be pondered. Having the data in many places versus one place means different choices in terms of solution architecture and programming. Spreading the data out can enable parallel processing techniques that overcome bottlenecks or pinch points. Pulling data together from multiple locations, multiple servers, or database management systems will be burdened by some amount of latency. That is the distance and time it takes to bring the data from its source(s) to the target. How much latency is introduced depends on the solution configuration and design of the database. Collocating data or information that is related or used in tandem is likely the only way to eliminate or severely limit the delay.
One area that fundamentally demands distribution is business intelligence (BI). The diverging requirements of online transaction processing databases and of data warehouse and data mart databases are usually handled by building separate systems. Whether the actual BI system is comprised of a distributed database is again a matter of history, capability, limitations, and personal preference of the architect, or more likely, the vendor selling his or her particular solution. Now, this does not mean that extracting data from the production database and "distributing" said data to Excel spreadsheets around the organization is a proper application of business intelligence; quite the contrary.
At the end of the day, there are architectures, designs, and techniques to meet, dare we say, any set of data requirements. The solution to choose is the one that meets your requirements and allows you to see a return on your investment sooner rather than later, while maintaining a low cost of ownership.