BizOps for the Data Scientist

In the market today, many topology approaches are closed in nature, bound by specific technological approaches and linear models. By contrast, advanced AI platforms should employ an open, source-agnostic approach. The platform’s architecture should deliver flexibility in several key ways:

• Data source extensibility. The platform’s architecture should not bound by any specific product, but feature an open data lake, algorithms, and more. Customers should be able to easily accommodate new data sources, including those from a wide range of third-party vendors.
• Architecture extensibility. To offer maximum long-term flexibility, a platform should enable partners and customers to introduce entirely new ontologies, without having to make any architectural changes.
• Ontology extensibility. Teams should be able to add different properties onto existing ontologies, and so easily accommodate organization-specific information, including tribal knowledge, naming or classification approaches, and so on.
• Robot extensibility. The ideal architecture should efficiently accommodate new robots as needed, while at the same time, enabling each robot to be employed against a unified, consistent data set.

In development, vendors can choose whether to create big data models that are very general or very specific. While very specific models may be pragmatic in the near term, they typically require manual modifications to accommodate new technologies, data types, and so on. Over time, these specific models tend to become brittle, leaving teams unable to react quickly to changes.

Instead, it is important to look for a platform that employs a property-graph based approach that has history awareness. With property graphs, complex relational lookups can be done instantaneously. As a result, they provide an excellent structure for doing ontological inference. By comparison, using a traditional relational database management system (RDBMS) for this model would require an impractical amount of join queries between schemas and tables, introducing an unacceptable level of performance-degrading latency.

In addition, here are a couple other characteristics to look for:

• Properties. The optimal machine-learning model uses properties to describe the qualitative specifics of a given entity. A customer may have monitoring agents that discover some properties of an element, such as the manufacturer, firmware version, software type, and so on. In addition, teams can leverage other data sources to further decorate the element with additional facts.
• Time journaling. Within the AI platform, all events should be captured in a time journal. These time-stamped records represent an immutable data point, and a valuable way to establish incremental observation of an environment. Through this approach, users and algorithmic analysts can observe structures and cross-domain interdependencies. For example, this enables a team responsible for networks to also see how changes in the broader environment may affect specific devices, such as when an application update causes traffic to start being sent to a different router, causing an unexpected workload spike.

These insights are drawn from our blog post, Putting Machine Learning Algorithms to Work, which includes more information on the optimal architecture for maximizing the value of data science.