I was recently asked about my opinion for the potential of Hadoop replacing a company’s data warehouse (DW). While there’s lots to be excited about when it comes to Hadoop, I’m not currently in the camp of folks that believe it’s practical to use Hadoop to replace a company’s DW. Most corporate DW systems are based on commercial relational database products and can store and manage multiple terabytes of data and support hundreds (if not thousands) of concurrent users. It’s fairly common for these systems to handle complex, mixed workloads –queries processing billions of rows across numerous tables along with simple primary key retrieval requests all while continually loading data. The challenge today is that Hadoop simply isn’t ready for this level of complexity.
All that being said, I do believe there’s a huge opportunity to use Hadoop to replace a significant amount of processing that is currently being handled by most DWs. Oh, and data warehouse user won’t be affected at all.
Let’s review a few fundamental details about the DW. There’s two basic data processing activities that occur on a DW: query processing and transformation processing. Query processing is servicing the SQL that’s submitted from all of the tools and applications on the users’ desktops, tablets, and phones. Transformation processing is the workload involved with converting data from their source application formats to the format required by the data warehouse. While the most visible activity to business users is query processing, it is typically the smaller of the two. Extracting and transforming the dozens (or hundreds) of source data files for the DW is a huge processing activity. In fact, most DWs are not sized for query processing; they are sized for the daily transformation processing effort.
It’s important to realize that one of the most critical service level agreements (SLAs) of a DW is data delivery. Business users want their data first thing each morning. That means the DW has to be sized to deliver data reliably each and every business morning. Since most platforms are anticipated to have a 3+ year life expectancy, IT has to size the DW system based on the worst case data volume scenario for that entire period (end of quarter, end of year, holidays, etc.) This means the DW is sized to address a maximum load that may only occur a few times during that entire period.
This is where the opportunity for Hadoop seems pretty obvious. Hadoop is a parallel, scalable framework that handles distributed batch processing and large data volumes. It’s really a set of tools and technologies for developers, not end users. This is probably why so many ETL (extract, transformation, and load) product vendors have ported their products to execute within a Hadoop environment. It only makes sense to migrate processing from a specialized platform to commodity hardware. Why bog down and over invest in your DW platform if you can handle the heavy lifting of transformation processing on a less expensive platform?
Introducing a new system to your DW environment will inevitably create new work for your DW architects and developers. However, the benefits are likely to be significant. While some might view such an endeavor as a creative way to justify purchasing new hardware and installing Hadoop, the real reason is to extend the life of the data warehouse (and save your company a bunch of money by deferring a DW upgrade)
Not long ago, I was asked to review a client’s program initiative that was focused on constructing a new customer repository that would establish a single version of truth. The client was very excited about using Master Data Management (MDM) to deliver their new customer view. The problem statement was well thought out: their customer data is spread across 11 different systems; users and developers retrieve data from different sources; reports reflect conflicting details; and an enormous amount of manual effort is required to manage the data. The project’s benefits were also well thought out: increased data quality, improved reporting accuracy, and improved end user data access. And, (as you can probably imagine), the crowning objective of the project was going to be creating a Single View of the Customer. The program’s stakeholders had done a good job of communicating the details: they reviewed the existing business challenges, identified the goals and objectives, and even provided a summary of high-level requirements. They were going to house all of their customer data on an MDM hub. There was only one problem: they needed a customer data mart, not an MDM hub.
I hate the idea of discussing technical terms and details with either business or IT staff. It gets particularly uncomfortable when someone was misinformed about a new technology (and this happens all the time when vendors roll out new products to their sales force). I won’t count the number of times that I’ve seen projects implemented with the wrong technology, because the organization wanted to get a copy of the latest and greatest technical toy. A few of my colleagues and I used to call this the “bright shiny project syndrome”. While it’s perfectly acceptable to acquire a new technology to solve a problem, it can be a very expensive to purchase a technology and force fit a solution that it doesn’t easily address.
It’s frequent that folks confuse the function and purpose of Master Data Management with Data Warehousing. I suspect the core of the problem is that when folks hear about the idea of “reference data” or a “golden record”, they have this mental picture of a single platform containing all of the data. While I can’t argue with the benefit of having all the data in one place (data warehousing has been around for more than 20 years), that’s not what MDM is about. Data Warehousing became popular because of its success in storing a company’s historical data to support cross-functional (multi-subject area) analysis. MDM is different; it’s focused on reconciling and tracking a single subject area’s reference data across the multitude of systems that create that data. Some examples of a subject area include customer, product, and location.
If you look at the single biggest obstacle in data integration, it’s dealing with all of the complexity of merging data from different systems. It’s fairly common for different application systems to use different reference data (The CRM system, the Sales system, and the Billing system each use different values to identify a single customer). The only way to link data from these different systems is to compare the reference data (names, addresses, phone numbers, etc.) from each system with the hope that there are enough identical values in each to support the match. The problem with this approach is that it simply doesn’t work when a single individual may have multiple name variations, multiple addresses, and multiple phone numbers. The only reasonable solution is the use of advanced algorithms that are specially designed to support the processing and matching of specific subject area details. That’s the secret sauce of MDM – and that’s what’s contained within a commercial MDM product.
The MDM hub not only contains master records (the details identifying each individual subject area entry), it also contains a cross reference list of each individual subject area entry along with the linkage details to every other application system. And, it’s continually updated as the values change within each individual system. The idea is that an MDM hub is a high performance, transactional system focused on matching and reconciling subject area reference data. While we’ve illustrated how this capability simplifies data warehouse development, this transactional capability also enables individual application systems to move and integrate data between transactional systems more efficiently too.
The enormous breadth and depth of corporate data makes it impractical to store all of our data within a single system. It’s become common practice to prune and trim the contents of our data warehouses to limit the breadth and history of data. If you consider recent advances with big data, cloud computing, and SaaS, it becomes even more apparent that storing all of a company’s subject area data in a single place isn’t practical. That’s one of the reasons that most companies have numerous data marts and operational applications integrating and loading their own data to support their highly diverse and unique business needs. An MDM hub is focused on tracking specific subject area details across multiple systems to allow anyone to find, gather, and integrate the data they need from any system.
I recently crossed paths with the above mentioned client. Their project was wildly successful – they ended up deploying both an MDM hub and a customer data mart to address their needs. They mentioned that one of the “aha” moments that occurred during our conversation was when they realized that they needed to refocus everyone’s attention towards the business value and benefits of the project instead of the details and functions of MDM. While I was thrilled with their program’s success, I was even more excited to learn that someone was finally able to compete against the “bright shiny project syndrome” and win.
IT organizations have spent enormous sums of money over the past 10-15 years attacking productivity. They’ve acquiring data integration tools, implemented improved development methodologies, and even reengineered requirements gathering methods to ensure business priority alignment. And the result of all of this investment? Today’s data integration developers are easily 10x to 20x more productive than the COBOL programmers of the past. This shouldn’t be a surprise to anyone – writing, compiling, linking, and testing 3rd generation code is much slower than today’s GUI-based, drag-and-drop development tools. The tools work; developers are faster, quicker, and better.
So, why does it still seem to take an eternity and cost a fortune to acquire and integrate new data into an existing report? The bottleneck has moved upstream: finding and extracting source data is complicated and time consuming. We’ve invested in our Integration Competency Centers to create an assembly line to streamline the process of transforming and converting data that is loaded into databases or applications. Unfortunately, we’ve not devoted any effort in simplifying access or understanding the actual raw source data that feeds the assembly line.
Henry Ford didn’t invent the assembly line, he revolutionized it. One of the changes that he introduced to the assembly line was simplifying and standardizing parts and the actual assembly process. Prior to Ford’s assembly line, car assembly was a custom effort that required highly trained craftsmen to shape, tool, and fit parts by hand (in a very time consuming process). The parts weren’t always uniform, so the craftsmen had to spend a significant amount of time fitting the parts together.
In most IT environments, source system access and data content varies across the different application systems dramatically. This forces developers to become data craftsmen in order to deal with the data idiosyncrasies associated with the numerous source systems common to most companies. Every system stores data in a custom and unique manner; it takes a lot of time to search and analyze source system data in order to identify the necessary content. (A popular ERP package stores its details in more than 10,000 tables) So, each new request often requires developers to create “from scratch” code to access and manipulate new data from a source system. If you dig a bit, you’ll probably find that many of your application systems generate dozens or hundreds (yes, hundreds) of custom extracts to deliver data to support the various production business needs within your company.
While most folks might think that custom extracts are a reasonably decent solution, they’re not. In fact, they’re a problem that will only get worse with time. (Remember, every extract requires development time and ongoing support.) You’ll be better off consolidating all of those extracts into a single set that includes all of the data. This will reduce processing time, reduce storage, reduce maintenance, and ultimately save a lot of money. You’ll have to spend some time designing and building these new extracts and getting folks to migrate to using them, but the benefits will be significant. (One of my clients was able to defer a platform upgrade due to the CPU and storage reduction caused by the consolidation and removal of all of the custom extracts).
Standardizing source data to reduce the data craftsmen problem isn’t rocket science, but it’s more than simply creating a data dump or generating a backup file. You need to deliver data in a manner that can be quickly and easily consumed by other systems. This means that the content needs to be reformatted from the unique (sometimes indecipherable) format of the host application into a format that everyone else can use. This can be easily addressed by delivering data into database tables or flat files (I know one client that delivers data in tab delimited spreadsheet format). The data should reflect the values generated by the source system in a format that everyone can understand – the content shouldn’t be modified for cleansed (this is source data, not content ready for business consumption). Delivery should occur in a frequent and regular basis along with a plan for archiving a decent amount of history.
This isn’t a new concept; this was a common approach in the days when custom coded IBM mainframe applications were all the rage. Back then, data sharing was a priority and every application generated standard extracts to reduce I/O and storage costs. There was also an extreme sensitivity to developer time. Requesting a custom extract was frowned upon and rarely approved. Finding and accessing the data was as simple as referencing the extract files that were made available from every application system.
When it comes to improving the delivery speed of new data to business users, maybe we can learn something from Henry Ford and the world of mainframe development.
Unless you’ve been hiding in a cave in the past year, you’ve probably heard of CEP (Complex Event Processing) or data stream analysis. Because a lot of real-time analysis focuses on discrete data elements rather than data sets, this technology allows users to query and manipulate discrete pieces of information, like events and messages, in real-time—without being encumbered by a traditional database management system.
The analogy here is that if you can’t bring Mohammed to the mountain, bring the mountain to Mohammed: why bother loading data into a database with a bunch of other records when I only need to manipulate a single record? Furthermore, this lets me analyze the data right after its time of creation! Since one of the biggest obstacles to query performance is disk I/O, why not bypass the I/O problem altogether?
I’m not challenging data warehousing and historical analysis. But the time has come to apply complex analytics and data manipulation against discrete records more efficiently. Some of the more common applications of this technology include fraud/transaction approval, event pattern recognition, and brokerage trading systems.
When it comes to ETL (Extract, Transform, and Load) processing, particularly in a real-time or so-called “trickle-feed” environment, CEP may actually provide a better approach to traditional ETL. CEP provides complex data manipulation directly against the individual record. There is no intermediary database. The architecture is inherently storage-efficient: if a second, third, or fourth application needs access to a particular data element, it doesn’t get its own copy. Instead, each application applies its own process. This prevents the unnecessary or reckless copying of source application content.
There are many industries need a real-time view of customer activities. For instance in the gaming industry when a customer inserts her card into a slot machine, the casino wants to provide a custom offer. Using traditional data warehouse technology, a significant amount of processing is required to capture the data, to transform and standardize it, to load it into a table, only to make it available to a query to identify the best offer. In the world of CEP we’d simply query the initial message and make the best offer.
Many ETL tools already use query language constructs and operators to manipulate data. They typically require the data to be loaded into a database. The major vendors have evolved to an “ELT” architecture: to leverage the underlying database engine to address performance. Why not simply tackle the performance problem directly and bypass the database altogether?
The promise of CEP a new set of business applications and capabilities. I’m also starting to believe that CEP could actually replace traditional ETL tools as a higher performance and easier-to-use alternative. The interesting part will be seeing how long before companies emerge from their caves and adopt it.
photo by Orin Zebest via Flickr (Creative Commons license)
As I wrote in last week’s blog post, a data warehouse appliance simplifies platform and system resource administration. It doesn’t simplify the traditional time-intensive efforts of managing and integrating disparate data and addressing performance and tuning of various applications that contend for the same resources.
Many data warehouse appliance vendors offer sophisticated parallel processing environments, query optimization, and specialized storage structures to improve query processing (e.g., columnar-based engines). It’s naïve to think that taking data from an SMP (Symmetric Multi-Processing) relational database and moving it into a parallel processing environment will effectively scale without any adjustments or changes. Moving onto an appliance can be likened to moving into a new house. When you move into a new, larger house, you quickly learn that it’s not as simple as dumping all of your stuff into the new house. The different dimensions of the new rooms cause you realize that some of your old furniture or rugs simple don’t fit. You inevitably have to make adjustments if you want to truly enjoy your new home. The same goes with a data warehouse appliance; it likely has numerous features to support growth and scalability; you have to make adjustments to leverage their benefits.
Companies that expect to simply dump their data from a few legacy data marts over to a new appliance should expect to confront some adjustments or their likely to experience some unpleasant surprises. Here are some that we’ve already seen.
Everyone agrees that the biggest cost issue behind building a data warehouse is ETL design and development. Hoping to migrate existing ETL jobs into a new hardware and processing environment without expecting rework is short-sighted. While you can probably force fit your existing job streams, you’ll inevitably misuse the new system, waste system resources, and dramatically reduce the lifespan of the appliance. Each appliance has its own way of handling the intensive resource requirements of data loading – in much the same way that each incumbent database product addresses these same situations. If you’ve justified an appliance through the benefits of consolidating multiple data marts (that contain duplicate data), it only makes sense to consolidate and integrate the ETL processes to prevent processing duplication and waste.
To assume that because you’ve built your ETL architecture leveraging the latest and greatest ETL software technology that you won’t have to review the underlying ETL architecture is also misguided. While there’s no question that migrating tool-based ETL jobs to a new platform can be much easier than lower-level code, the issue at hand isn’t the source and destination– it’s the underlying table structures. Not every table will change in definition on a new platform, but the largest (and most used) table content is the most likely candidate for review and redesign. Each appliance handles data distribution and database design differently. Consequently, since the underlying table structures are likely to require adjustment, plan on a redesign of the actual ETL process too.
I’m also surprised by the casual attitude regarding technical training. After all, it’s just a SQL database, right? But application developers and data warehouse development staff need to understand the differences of the appliance product (after all, it’s a different database version or product). While most of this knowledge can be gained through reading the manuals – when was the last time the DBAs or database developers actually had a full-set of manuals—much less the time required to read them? The investment in training isn’t significant—usually just a few days of classes. If you’re going to provide your developers with a product that claims to bigger, better, and faster than its competitors, doesn’t it make sense to prepare them adequately to use it?
There’s also an assumption that—since most data warehouse appliance vendors are software-only—that there are no hardware implications. On the contrary, you should expect to change your existing hardware. The way memory and storage are configured on a data warehouse appliance can differ from a general-purpose server, but it’s still rare that the hardware costs are factored into the development plan. And believing that older servers can be re-purposed has turned out to be a myth. If you ‘re attempting to support more storage, more processing, and more users, how can using older equipment (with the related higher maintenance costs) make financial sense?
You could certainly fork-lift your data, leave all the ETL jobs alone, and not change any processing. Then again, you could save a fortune on a new data warehouse appliance and simply do nothing. After all, no one argues with the savings associated with doing nothing—except, of course, the users that need the data to run your business.
photo by Bien Stephenson via Flickr (Creative Commons License)
In the motion picture industry, studios separate responsibilities for creating content from responsibilities for distributing content. The people who make the movies option the scripts, hire the talent, and film the scenes. The distributors of the films, on the other hand, figure out how to package and deploy the films. They need to know which theaters require 30 millimeter versus 70 millimeter formats, or even IMAX. They also deal with DVD packaging, including different international DVD formats. The industry understands the importance of having a supply chain that differentiates between the roles of content creation, content packaging, and distribution.
In IT we’re very quick to point to our operational systems as creators and owners of data. But maybe the solution is that IT establishes a functional team that’s responsible for data packaging and distribution, just like the movie industry.
Traditionally data formats and standards have fallen into the realm of the architecture team. Unfortunately this is typically a paper-only activity without teeth. A data distribution team wouldn’t focus on paperwork. They would be focused on data logistics, receiving content from the various source systems and packaging the data for consumption by other systems. This isn’t about implementing a specific platform to store or move data. It’s about active management of corporate data content.
One of the biggest development challenges is the hunting expedition that developers go on to find and acquire the data they need. Most aren’t aware of all their choices, let alone the optimal systems of record.
Currently every application, data mart, data warehouse, reporting system that needs data from another system follows a specific set of procedures to obtain that data. Each system requests different data formats, different delivery schedules, and different content. Everything is custom, there are few if any standards, and there are no economies of scale.
This will also unburden the various application teams from building and maintaining the never ending volume of custom extract requests. The only way to stop the madness is to compartmentalize content creation from data packaging and distribution. This means establishing a data supply chain that separates data creators from data distribution from consumers. Who knew IT infrastructure was just like the movies?