Oracle’s Machine Learning Strategy
It’s hardly news that technology companies are working with machine learning. It’s also hardly news that technology companies targeting enterprises are working AI into their solutions. But when Oracle invited a small group of analysts to peek into the company’s internal machine learning summit, it was the first step toward letting the world in on its own secrets.
Although Oracle has been actively supported data mining, machine learning’s (ML’s) predecessor, in the database since the 9i generation nearly 20 years ago. it has not been as closely associated with AI as most of its key rivals. A big reason for that is that many of them are also consumer brands. We’re talking about Amazon, Google, and Microsoft, and if you want to widen the net outside Oracle’s space, Facebook and Netflix. Their recommendation and content curation engines touch people directly every day.
Because it caters to enterprises, IBM is the more apt comparison for Oracle. But here again, there’s a big difference. IBM staked the future of its company on a fanciful AI concept and brand – cognitive computing and Watson — and appeared on Jeopardy to drive the point home. IBM may have gained some AI mindshare after that. But after beating Ken Jennings and Brad Rutter, it saw the limitations of its vision. Cognitive computing might have been ready for prime time TV, but not for prime time. Fast forward to the present, and the Watsonbrand has shifted to mean all things AI. Meanwhile, Oracle didn’t have to do any pivoting on AI because, well, it really wasn’t talking about it until now.
Until recently, Oracle’s AI positioning was centered around its Autonomous Database, Adaptive Intelligence Applications (AI apps), and the advanced analytics option for its database (where it was one of the first providers to offer a scale-out in-database deployment of R). These offerings had meaning for business, but not to the public at large. Excluding those offerings, ML was not positioned as a product feature. And while a number of analytics software players are spotlighting AI with collaboration and model deployment tools catering to data scientists, Oracle is only now actively ramping up engagement with data scientists.
But Oracle is a tech company, and although it hasn’t trumpeted it, has been working with machine learning, and its precursor, data mining, since the turn of the millennium. Much of that work was providing capabilities for model building inside the database. But more of that work was under the hood, initially with rules engines, and then with machine learning, to improve the performance of the database. This work began many years prior to the introduction of the autonomous database, which uses machine learning, not only to improve its performance, but run the database.
It started with a precursor, using rules-based expert systems to troubleshoot and help DBAs remediate issues with performance and availability. Starting with Oracle Database 12.2c, released roughly five years ago, it began employing ML models to take advantage of their flexibility. Oracle conducted a supervised learning approach on an Oracle RAC cluster in the lab to build, train, and deploy the models using a combination of log data and metrics harvested from its SaaS business (model development was not done on live, production databases).
The exercise began by establishing the baseline of what patterns of metrics and logs represented normal and subpar operation. For abnormal operation, upstream events and metrics were analyzed to build early diagnostic capabilities into the models, and downstream data on which to build predictives. The models used Bayesian approaches that factor in performance history to detect the probability of performance drop or outage. On the horizon, Oracle will extend these models to account for the specific configuration of the database, and it is also working on models that diagnose database wait states (it is currently extending them to also cover the storage subsystem).
Don’t confuse these internal ML projects with making the database autonomous (or self-running). They were modestly aimed at improving the ability to diagnose runtime issues as opposed to taking over and running the database. But they were among the contributing pieces, along with other features like automatic storage management, automatic standby management, or statistics gathering, that were steppingstones to making the autonomous database possible. And by the way, while Oracle is currently the only provider to offer a completely self-running database cloud service, we don’t believe that they are alone in employing ML to tweak operations under the hood.
But now Oracle is coming out of its shell to promote and expand the portfolio of products, including applications and tools, that are either built around ML or use it prominently as a feature. Oracle is taking a dual-track strategy of embedding ML into its products that are designed to reach the mainstream of its customer base, and data science-oriented tools and frameworks for the more sophisticated customers who have teams of data scientists and require tools to crank out their own custom solutions. The common thread is that the majority of these offerings will be delivered as managed services in its public cloud.
Beyond the Autonomous Database, which one of the managed services that run in the Oracle Public Cloud, there are the Adaptive Intelligent Applications (AI Apps) that embed ML into ERP, customer engagement, manufacturing, and HCM. Oracle is only at the beginning of this journey here, counting 19 such modules. We expect that, just as analytics are being embedded into modern enterprise applications, that ML-enhanced processes will become the norm. But for now, Oracle segments them out as discrete products. Here again, Oracle is hardly alone. SAP is also at the beginning of its journey, with one of the goals of the Leonardo initiative being to identify use cases for applying AI to enhance core business processes.
At Oracle, other examples include DataFox, a recent acquisition, which is applying natural language processing to build, in essence, a next generation answer to Dun & Bradstreet company listings that profile prospective B2B partners and customers; Oracle Digital Assistant, which provides chatbots; and Fusion Financials, which perform smart document recognition for invoicing and expenses.
Additionally, Oracle is also utilizing its graph data store (part of the Oracle Spatial and Graph option for Oracle Database) for various risk management and security related use cases, such as detecting and analyzing cyberthreats or incidences of money laundering; here, graph data store and ML models feed on each other, as graph represents the complex webs of relationships, on which ML models can identify suspicious activity hotspots, and in turn, uncover hidden relationships or interdependencies.
With the DataScience.com acquisition, Oracle extends its collaboration platform for data scientists to build and manage ML models with a managed service available from the Oracle Public Cloud. The offering, now called Oracle Cloud Infrastructure Data Science is a general-purpose data science platform that works with a variety of data sources (not only Oracle databases) and supports use of open source frameworks and libraries along with Oracle’s proprietary Python libraries. It allows data scientists to explore data, train, save, and deploy models; we expect that it will subsequently add model lifecycle management. For development and deployment, Oracle offers an Apache Zeppelin-based notebook. At run time, these models can execute on any target, but the Oracle Database supports running models inside the database through SQL, R, and soon, Python APIs.