BI: Real time or right time?
|These days, applying to college is much like applying for a credit card. Both processes impact people’s lives and, increasingly, both processes require extensive demographic modeling to rank prospects. Building those models behind the scenes requires serious data mining and analysis of hundreds of thousands or millions of records to identify those prospects most likely to earn straight-A grades or pay their credit card bills on time.|
Demographic models may take hours or days to compute, but once built, they can judge an applicant’s qualifications on the spot. In short, these processes combine traditional offline analyses associated with business intelligence (BI) problems with carefully bounded real-time use.
Noel-Levitz is a higher education consulting firm that helps colleges and universities to sort out prospects during the admissions process, and demographic models are at the core of its business. Institutions subscribing to Noel-Levitz’s service initially have their lists screened for likely prospects. They are then re-screened once real applications are received, providing institutions with detailed analysis from which they can make their admissions decisions.
From a timing standpoint, that is when things get dicey, according to Tim Thein, senior vice president at Noel-Levitz, which has offices in Iowa City, Iowa, Denver and Guelph, Ontario. “There is a five- to six-day overall lifetime for any response,” he explained. Because most applicants are likely weighing offers from several institutions, those fabled acceptance letters must be mailed as quickly as possible. “At that point, a one-week delay in response is not acceptable,” he said.
To accelerate that final step, Noel-Levitz is phasing in a new Web portal that will enable subscribers to submit lists of applicants via SOAP message-based Web services requests. Using middleware and data mining tools from SAS Institute, Noel-Levitz is able to process up to 80,000 names in 15 minutes, or a single name almost instantly.
However, whether the interactive applications of Noel-Levitz -- or credit card issuers for that matter -- actually constitute real-time business intelligence is debatable, since most of the heavy lifting continues to be performed offline. What passes for real-time analysis is a relatively modest check against an existing scoring model. But new network-based approaches that intercept events, transactions or messages coming over the wire, apply relatively simple analytics, and then alert the right recipients at the right time might transform real-time BI from hype to reality.
What’s hot? What’s not?
There is little question that businesses can benefit from tools that help them to make decisions more rapidly. Regardless of whether it entails optimizing a supply chain or balancing trading positions, gaining real-time visibility may have huge benefits. But whether that demands real-time analysis may be another story.
In each case, the development manager must navigate a landscape of hyperbole. It is difficult to debate the need for real-time analytics because the definition of real time itself can be so highly subjective. Furthermore, when it comes to real time, there is a question of whether you can have too much of a good thing. For instance, if the information is so instantaneous, could business analysts find themselves reacting to momentary spikes that obscure the big picture?
The debates over real-time business intelligence have been fueled in large part by vendor marketing slogans promoting concepts such as “zero latency,” “on-demand” integration, “extreme analytics” and, of course, the granddaddy of them all, “The Real-Time Enterprise.” Arguably, the emergence of Business Activity Monitoring (BAM) solutions that enable enterprises to fix problems while they are occurring has further ratcheted up the debate.
Maybe there is a middle ground. Some observers suggest augmenting carefully scheduled BI “batches” with event-driven capabilities as a necessary step. These applications could be called “right-time” systems because different portions of the analyses are performed when practical and as needed.
Call center on hold
Call centers are ready examples of the distinctions to be made between analysis and visibility. Because customers do not like to be kept on hold, merchants who operate call centers obviously prefer optimizing staffing ratios to maintain desired service levels without incurring excessive labor costs. At PNC Bank’s call center, where operators are scheduled in 15-minute shifts, the bank cannot afford to wait until the next shift or the following day to analyze traffic loads. Currently, the bank monitors traffic levels live, in real time, from a Lucent phone switch and uses Information Builders’ WebFocus reports, updated hourly, to couch its staffing decisions.
Would it make sense for PNC to update its WebFocus reports with the same real-time data off the Lucent switch? Although there is a technical hurdle -- the call center’s system does not have enough connections with the Lucent switch to support more frequent downloads -- that problem is solvable. More important is the question of what the bank would do with all that data. Conceivably, if the analyses changed from minute to minute, managers could find themselves overreacting to momentary spikes in traffic. “It’s not enough to know where we are right now,” said Aaron Leaman, project manager. “We need to know where we’ve been for the last two or three hours, or how we compare to the same time yesterday.”
What about the supply chain? The emergence of just-in-time manufacturing has prompted popular images of trucks pulling up to loading docks that are scheduled by the hour. Although strictly timed, these movements are choreographed hours or days in advance because it takes time for trucks to travel the roads to make deliveries. Furthermore, the analyses used for judging the effectiveness of inventory deployment and logistics are typically run well after the fact, when the details can be more readily digested.
According to Thomas Bornemann, managing director of consumer products at Durham, N.C.-based Clarkston Consulting, when industry leaders, such as Wal-Mart, insist on keeping retail shelves constantly filled, that could change the notion of how “hot” supply-chain business intelligence must be.
Bornemann cited a reverse-auction process from one of his firm’s clients that involves a combination of Web-based, real-time analytics and manual decision-making. The manufacturer publishes its supply-chain forecasts to downstream business partners, who package the product for delivery to retailers. Based on a variety of factors, including historical analysis of price vs. demand and current cost data, the manufacturer calculates and publishes to the Web the price it will pay for packaging services. The analytic component then quickly analyzes supplier responses and dispatches the results to a supervisor, who subsequently decides whether to increase the price if the co-packer response is not adequate. Although the last step of the process remains manual, Bornemann claims that even this function could be automated with yet another real-time analysis if smart agent technology were used.
The BAM factor
Emerging BAM tools provide displays that are often described as “dashboards” that summarize how well a company is executing based on key performance indicators (KPIs). For instance, a financial services firm’s dashboard might show the profitability of its various product lines, a retailer’s dashboard might cover pricing and promotions, while a manufacturer’s dashboard could display supply-chain performance.
In effect, BAM tools supplement the content-based routing capabilities of EAI brokers. For instance, while an EAI system routes an inventory planning transaction from a J.D. Edwards ERP system to a mainframe-based procurement system, the BAM tool could inspect the inventory message itself to see if inventory is about to exceed a certain threshold and trigger an alert.
BAM dashboards are not necessarily analytic applications, since the alerts simply function as wake-up calls that explain what is happening, rather than why. However, logic could be built into BAM alerts that also trigger spot analyses comparing current trends to recent history. For instance, Tibco’s BusinessFactor application provides a Java-based tool that provides some on-the-fly analytical drill-down capabilities from a BAM dashboard. Similarly, webMethods provides a dashboard, backing it with links into the vendor’s EAI framework and Informatica’s data integration middleware. A major caveat, however, is that the data on a BAM dashboard may -- or may not -- be real time.
Scott Fingerhut, senior product marketing manager for Tibco’s BusinessFactor product line, described a scenario with a health-care client that is analyzing movements to and from the emergency room. When patients wait too long at any given stage of the process, it could trigger an analysis to see whether, for instance, a 21-minute wait at 11:00 on Friday night is unreasonable based on historical data. He hesitates to term this real-time analytics, however. “It’s more benchmarking against expectations,” he said, noting that “analytics involves performing mathematical equations to see how we performed over a period of time.”
It’s the network, stupid
Traditionally, BI systems have relied on data warehouses or operational data stores that provide data that is kept separate from the transaction system. A recent refinement of the traditional BI architecture has been the introduction of trickle feeds that supplement or replace batch downloads that populate data warehouses. Conceivably, these trickle feeds could enable BI systems to work with more current data. While having current data in the data warehouse is not the same as running a real-time analysis, these systems could conceivably be enhanced with alerting capabilities whereby smart agents would monitor data flows and dispatch alerts when anomalies are detected. A further refinement could involve the convergence of data warehouse and enterprise integration infrastructures by enabling data warehouses to tap into the current flow of transactions or messages directed by EAI brokers.
It is fair to ask whether turbocharging the process for updating data warehouses is necessarily the most efficient route to real-time analysis, since the process still involves a sequence of multiple events before an analysis can be generated. According to database pioneer Jnan Dash, who was active in the development of DB2 and Oracle, the answer is “no.” “We need nimble, lightweight, zero-install browsers that filter events, grab them and notify us in real time,” he contends.
Dash’s vision implies an approach that is similar to that of an industrial process control system. There, sensors or probes monitor delicate processes to ensure that they are operating within acceptable safety margins, with current data maintained in memory-resident databases. When the process begins to drift off spec -- a parameter that is detected directly or computed as part of a trend analysis -- alarms are activated that alert operator or supervisor attention, paving the way for the process to be repaired manually or automatically, or even taken offline.
Stealing a page from publish/subscribe models, several novel approaches are emerging that, in effect, treat networks as event busses, while providing “listeners” that perform their own form of selective routing. These “listeners” could be tuned to sources such as transaction systems, EAI hubs, BAM dashboards or specific Web portals. As with BAM systems, alerts are not necessarily synonymous with analysis, but triggering logic could be built in to provide the necessary intelligence, using tools from the usual BI vendor suspects as well as from emerging players.
For instance, emerging middleware provider Bristol Technology Inc. is adding transaction tracking to the usual BAM palette. Compared to systems management frameworks, which detect the overall health of a transaction system, or a transaction monitor, which optimizes the processing transactions within a single database, Bristol’s tools interrogate messages sent via JMS or MQSeries to trace all the points where a transaction travels, check the inputs to detect which systems were called, and tally what values are passed back.
Conversely, Kabira Technologies Inc., a provider of message-switching systems for Unix environments, specializes in high-volume, in-memory transaction systems that are commonplace throughout the financial services and telco sectors. Having developed niche solutions for processes such as fraud detection and network provisioning, the Kabira message switch could be adapted for analytic purposes. For instance, a telco customer uses the system for customer satisfaction analyses to reduce churn rates. During peak traffic conditions, the switch “listens” for incidences of dropped calls; when the figure breaches a certain threshold, the system transmits text messages to subscribers offering them free phone time as compensation.
Apama, a new provider of real-time data mining solutions, offers an alternative to the in-memory databases that have been common in sectors like financial services and telco. “The problems with those approaches is that, because they are databases, you have to index everything,” explained Steve Allison, director of product marketing in the firm’s New York offices, noting that indexing operations add their own time lags. Furthermore, there are significant scalability limitations regarding placing rules into memory because every piece of data must be run through the rule.
Instead, Apama streams data “through” a query that scans for patterns, and discards extraneous data falling outside the pattern. Allison describes a financial arbitrage scenario that scans for pricing differences across multiple streams of data, calculates a volume-weighted average price, and then finds buy or sell opportunities based on the derived figure. Allison also claims that the same approach could apply to BAM dashboards, where companies compare ERP forecasts for inventory, production schedules and sales to optimize capital flows, and trigger analyses when the deviation of actual numbers from forecast exceeds certain thresholds.
Right time, not spike time
In the world of factory automation, real-time refers to the type of “deterministic” response that is necessary when running a machine. In the computing world, real time is a relative concept describing the ability to interactively access or update data that is current, rather than data that was last refreshed during a previous batch run or reporting period. Consequently, the concept of “Real-Time Business Intelligence” may prove too subjective to define or debate.
Henry Morris, vice president, applications and information access at IDC, prefers another term, “event-driven,” which he defines as business intelligence that is conducted not at the convenience of IT, but at the request of the end user. In the supply-chain example, the analysis is performed while the process is occurring, but not necessarily down to the split second or minute.
Perhaps a better way to describe these BI applications is “right time.” For instance, credit-scoring applications may perform 99% of the analysis offline, saving the very last calculation for when the actual decision is requested. “A lot of BI continues to be conducted on a scheduled basis,” noted Morris, “but you may need to augment that with event-based capabilities.”
Either way, if business intelligence is produced on the spot, while a process is still happening, it will probably be different from traditional analytic reports that are generated after the fact. In all likelihood, it will be a subset of what is typically performed offline because people and machines have finite capacities. “It’s not a case where you will have 50,000 metrics that flash on your desktop in real time,” said Sanjay Poonen, senior vice president of worldwide marketing at Informatica Corp., Redwood City, Calif.
Please see the following related stories:
“Real-time light, real-time anywhere” by Alan Earls
“What to look for in a BI architecture” by Wayne Eckerson
“What’s real at Fleet?” by Jack Vaughan