July 19, 2021
We are increasingly surrounded by intelligent IoT devices, which have become an essential part of our lives and an integral component of business and industrial infrastructures. Smart watches report biometrics like blood pressure and heartrate; sensor hubs on long-haul trucks and delivery vehicles report telemetry about location, engine and cargo health, and driver behavior; sensors in smart cities report traffic flow and unusual sounds; card-key access devices in companies track entries and exits within businesses and factories; cyber agents probe for unusual behavior in large network infrastructures. The list goes on.
How are we managing the torrent of telemetry that flows into analytics systems from these devices? Today’s streaming analytics architectures are not equipped to make sense of this rapidly changing information and react to it as it arrives. The best they can usually do in real-time using general purpose tools is to filter and look for patterns of interest. The heavy lifting is deferred to the back office. The following diagram illustrates a typical workflow. Incoming data is saved into data storage (historian database or log store) for query by operational managers who must attempt to find the highest priority issues that require their attention. This data is also periodically uploaded to a data lake for offline batch analysis that calculates key statistics and looks for big trends that can help optimize operations.
What’s missing in this picture? This architecture does not apply computing resources to track the myriad data sources sending telemetry and continuously look for issues and opportunities that need immediate responses. For example, if a health tracking device indicates that a specific person with known health condition and medications is likely to have an impending medical issue, this person needs to be alerted within seconds. If temperature-sensitive cargo in a long haul truck is about to be impacted by an erratic refrigeration system with known erratic behavior and repair history, the driver needs to be informed immediately. If a cyber network agent has observed an unusual pattern of failed login attempts, it needs to alert downstream network nodes (servers and routers) to block the kill chain in a potential attack.
To address these challenges and countless others like them, we need autonomous, deep introspection on incoming data as it arrives and immediate responses. The technology that can do this is called in-memory computing. What makes in-memory computing unique and powerful is its two-fold ability to host fast-changing data in memory and run analytics code within a few milliseconds after new data arrives. It can do this simultaneously for millions of devices. Unlike manual or automatic log queries, in-memory computing can continuously run analytics code on all incoming data and instantly find issues. And it can maintain contextual information about every data source (like the medical history of a device wearer or the maintenance history of a refrigeration system) and keep it immediately at hand to enhance the analysis. While offline, big data analytics can provide deep introspection, they produce answers in minutes or hours instead of milliseconds, so they can’t match the timeliness of in-memory computing on live data.
The following diagram illustrates the addition of real-time device tracking with in-memory computing to a conventional analytics system. Note that it runs alongside existing components. It adds the ability to continuously examine incoming telemetry and generate both feedback to the data sources (usually, devices) and alerts for personnel in milliseconds:
Let’s take a closer look at today’s conventional streaming analytics architectures, which can be hosted in the cloud or on-premises. As shown in the following diagram, a typical analytics system receives messages from a message hub, such as Kafka, which buffers incoming messages from the data sources until they can be processed. Most analytics systems have event dashboards and perform rudimentary real-time processing, which may include filtering an aggregated incoming message stream and extracting patterns of interest. These real-time components then deliver messages to data storage, which can include a historian database for logging and query and a data lake for offline, batch processing using big data tools such as Spark:
Conventional streaming analytics systems run either manual queries or automated, log-based queries to identify actionable events. Since big data analyses can take minutes or hours to run, they are typically used to look for big trends, like the fuel efficiency and on-time delivery rate of a trucking fleet, instead of emerging issues that need immediate attention. These limitations create an opportunity for real-time device tracking to fill the gap.
As shown in the following diagram, an in-memory computing system performing real-time device tracking can run alongside the other components of a conventional streaming analytics solution and provide autonomous introspection of the data streams from each device. Hosted on a cluster of physical or virtual servers, it maintains memory-based state information about the history and dynamically evolving state of every data source. As messages flow in, the in-memory compute cluster examines and analyzes them separately for each data source using application-defined analytics code. This code makes use of the device’s state information to help identify emerging issues and trigger alerts or feedback to the device. In-memory computing has the speed and scalability needed to generate responses within milliseconds, and it can evaluate and report aggregate trends every few seconds.
Because in-memory computing can store contextual data and process messages separately for each data source, it can organize application code using a software-based digital twin for each device, as illustrated in the diagram above. Instead of using the digital twin concept to model the inner workings of the device, a real-time digital twin tracks the device’s evolving state coupled with its parameters and history to detect and predict issues needing immediate attention. This provides an object-oriented mechanism that simplifies the construction of real-time application code that needs to evaluate incoming messages in the context of the device’s dynamic state. For example, it enables a medical application to determine the importance of a change in heart rate for a device wearer based on the individual’s current activity, age, medications, and medical history.
The complex web of communicating devices that surrounds us needs intelligent, real-time device tracking to extract its full benefits. Conventional streaming analytics architectures have not kept up with the growing demands of IoT. With its combination of fast data storage, low-latency processing and ease of use, in-memory computing can fill the gap while complementing the benefits provided by historian databases and data lakes. It can add the immediate feedback that IoT applications need and boost situational awareness to a new level, finally enabling IoT to deliver on its promises.
