Despite the looming economic recession, many companies aren’t pulling back on technology investments. In fact, according to a CNBC survey, 75 percent of tech leaders expect their technology spend to increase this year. But why?
Digital infrastructure agility and resiliency are particularly crucial for enterprises in periods of economic uncertainty. Digital technology can help enterprises cut costs by streamlining processes and automating routine tasks. According to Harvard Business Review research, during economic uncertainty, “companies should prioritize ‘self-funding’ transformation projects that pay off quickly, such as automating tasks or adopting data-driven decision-making.”
RFID-based applications have a lot to contribute here, but there is a need to get these kinds of data streams integrated more quickly and easily into mainstream business processes. Accelerated big data analytics application platforms can help, especially those which provide self-service analytic options. However, a key issue becomes planning appropriate capabilities, especially the support for scalable spatiotemporal (ST) operations. Even small numbers of RFID devices generate large datasets, so there is a business danger in piloting something which then cannot realistically be scaled to production. The use of standards-based geoSQL on a platform designed to handle large ST data is often critical to ensuring quick initial delivery, but also maintainability and extensibility.
Investing in Decision Intelligence
These days, companies face many challenges such as supply chain disruptions, inflation, geopolitical instability, the lingering effects of the pandemic and climate change—all leading to business volatility. Most companies suffer when the economy is down, primarily because demand and revenue fall and uncertainty about the future increases. To prepare for and weather a recession successfully, businesses need to be flexible and make key adjustments. The right technology investments enable increased operational agility for organizations and provide competitive advantage.
Michael Flaxman
There are two lower-level technology innovations to pay attention to here: columnar databases and GPUs. This is because we are at an inflection point for technology, where enterprises have access to considerably more data than ever before in history. To generate optimal business value, data collected from RFID tags almost always needs to be tabularly and spatially joined to that from Internet of Things (IoT) sensors, satellites, wireless GPS devices and more. In other words, big data is about data variety and velocity, not just volume.
Enterprise data storage is sufficiently advanced to store all this data as it is continuously collected. However, integrating such data quickly and on demand remains an important challenge. Big data has gravity, in the sense that moving it across a data center or between such centers can be costly in both time and treasure. So an important consideration becomes the ability to minimize such transfers by moving only the minimum data required to answer a query or set of queries.
Columnar databases and storage formats are an important advance here, especially for ST data. That is because with this type of technology, it becomes very efficient to push down queries to grab data bound in both space and time. For example, if your query is designed to determine if a particular RFID tag on a shipping container has left port yesterday, your business logic has already defined ST bounds. With federated push-down logic, you can avoid loading millions to billions of records which are older or relate to other geographies.
A second major innovation is the advent of analytical software leveraging graphics processing units (GPUs). While previously used mostly for gaming or crypto mining, GPUs now provide the computing power to process and analyze millions of data points within a matter of minutes, rather than weeks. This is not a small incremental change. Systems built to use GPUs efficiently are 10 to 100 times faster than legacy systems.
Leveraging GPUs opens the door for businesses to use big data insights and visualization to make smarter high-impact decisions based on the massive datasets at their fingertips, all within minutes or even seconds. This allows them to identify new revenue and cost savings opportunities quickly, which are key competitive advantages, particularly in the recessionary phase of a business cycle.
Challenges for Reaping the Benefits of Accelerated Data Analytics
So, if most of your RFID and business data is housed in conventional databases on CPU architectures, do you need to throw all this away and start over again? Fortunately not! Better strategies are available. In 2023, all databases already know how to talk to all other databases, and all business intelligence tools know how to talk to all databases. A better strategy here is thus to selectively add an “acceleration tier.” This tier doesn’t replace your core databases, but rather complements them in a specific set of highly demanding applications.
Extending the RFID use on shipping containers example above, imagine that the database of record is postGIS. That’s a great 40-plus-year-old database, now also available in the cloud. But while it can perform sophisticated geoSQL, doing so on large ST data volumes requires continuous data re-indexing and long batch queries. Since it has no intrinsic map rendering capabilities on CPU, much less GPU, postGIS needs to be paired with some kind of map rendering server or client. Unfortunately, that architecture essentially requires pushing very large geotemporal data around each time you want to drill down visually to find a particular container.
Instead of that awkward and non-performant last step, consider the acceleration tier strategy. Depending on your existing data location and security requirements, you could host the acceleration tier either on-premise or in the cloud. The critical thing there is simply that it has GPU plus fast access to the relevant data. Fortunately with modern container-based applications, installation in either context is pretty easy. If you don’t have an appropriate on-premise kit, you can even use the cloud vendors to trial various hardware options by the hour.
At the cost of running a docker or kubernetes command and maybe opening up a few ports, you’ll now have a modern and performant GPU-analytics tier. To test it out, you only have two further steps. First, you need to make and persist a connection to your legacy database, entering the appropriate credentials. Second and finally, you need to create a map-based dashboard. The details obviously vary by system, but with modern GPU analytics tools, map-based charts are typically built in. Et voila, 10 to 100 times drill-down RFID tracking and map rendering performance in a single sitting.
Analyzing IoT and RFID Data to Mine Business Insights
With that infrastructure bit in place, your organization will already be much better provisioned. You’ll have high-performance analytics on big data, and also likely a much-improved level of self-service analytics. Time to harvest some more business value! The dashboard already discussed would be an example of human-in-the-loop analytics. You would gain efficiency and business value, for example, because instead of an employee calling the shipping agent and being placed on hold, they could track an important container directly themselves. That might save 5 to 15 minutes for each occurrence.
But adding a second level of automation might further improve return on investment (ROI). For example, most people don’t care about operations going as expected. They want to know as soon as possible when things are going badly. In analytics, this is usually known as anomaly detection. So instead of having users log into a dashboard to track an RFID, how about adding a user-triggered change tracker? The user application shifts to an interface to specify the desired communications details. Does the user care if the container is ahead of schedule? Three hours behind? The compute architecture remains basically the same as prior, with the addition of a notification metric.
Again, the details will vary by system, but essentially the user interface just lets the user adjust how chatty they want their notifications to be, and by which means they want them delivered (email, text, slack, etc.). All of a sudden, however, every employee of the company has one less thing to worry about—unless they really need to worry about it. Anomaly detection is of high business value because it helps staff members focus on what’s really important.
Lastly, let’s consider a fancier case which is still pretty general. Imagine that you are responsible for a complex logistics chain to manufacture a widget. As is typical in logistics, you want to minimize your inventory and storage costs while still maintaining reliability in component delivery. If you are missing one essential component, your production line goes down. But if you have to maintain three months of inventory of everything, your inventory and storage costs are not competitive.
This is a good case for predictive analytics. In concept, we’re no longer worried about the location of a single RFID tag, or even the delay on a single tag, but rather predicting the consequences of multiple supply chain variations. One way of approaching this is to apply predictive analytics first to prediction of delays, and then to consequences of such delays.
Imagine we hand this task to our data science team. They will essentially build two coupled high-level models, one for delays and the other for consequences. Along the way, they’ll typically generate hundreds of test model instances in order to determine the best model architecture. Each time one of these is trained, they’ll want access to as much historical data as possible, so that for example they can calibrate to differences between suppliers or shipping routes.
While the process will undoubtedly involve a lot of new jargon about “feature engineering” and “model drift,” the underlying issue is that data science requires very heavy forms of data access, usually to full database history. If you do not plan for this, any data science group will likely overwhelm the resources of any conventional data architecture. For example, they will be hammering your production databases day and night. Fortunately, this is actually another area where GPU-based RFID analytics can be very efficient. Data science teams routinely use and manage GPUs.
So in this hypothetical case, you’d be well served to house at least one instance of your acceleration tier within your data science group. That way, the heavy query loads required in model training will mostly be isolated to the acceleration tier and away from production. As a secondary benefit, the results of these models will already be available in a business analytics tool along with the rest of your enterprise data. This means model testing and delivery times can be substantially reduced, saving both time and money.
Preparation and Intelligent Decision Making Are Key
As these examples hopefully illustrate, there are a variety of ways in which accelerated analytics can drive high business value, even in lean times. Business intelligence and data visualization turn the flood of RFID and IoT data into new insights and enable new ways of solving problems. The convergence of the IoT and big data promises tremendous new business value and opportunities for enterprises across all major industries. As enterprises grapple with the volume, velocity and variety of their data, they can potentially discover areas of improvements and cost savings, as well as uncover new revenue streams.
Studying the historical differences between companies that fare well during recessions and those that do not, Harvard Business Review references a Bain report that found the top factor that made a difference was preparation. These days, that preparation translates into investments in enterprise tech that supports digital transformation and provides significant ROI for the business. By deploying accelerated RFID analytics, enterprises will have the capability to proactively identify areas of improvement and focus resources to recession-proof the business in times of economic uncertainty.
Dr. Michael Flaxman is the VP of product at HEAVY.AI. His team is focussed on the combination of geographic analysis with machine learning. Michael has served on the faculties of MIT, Harvard and the University of Oregon, and he has participated in GIS projects in 17 countries. He has been a Fulbright fellow; has served as an advisor to the Interamerican Development Bank, the World Bank and the National Science Foundation; and served as industry manager for architecture, engineering and construction at ESRI. Michael received his doctorate in design from Harvard in 2001, and he holds a master’s degree in community and regional planning from the University of Oregon and a bachelor’s degree in biology from Reed College.