SAS

Perhaps the biggest differentiator for SAS in this space is its analytic capabilities, though it’s also notable for its performance. The options available are illustrated in Figure 2, many of which have only been introduced in the latest release (5.1). 
A key feature is the continuous improvement of in-stream models, using machine learning.

Figure 2

As far as Figure 2 is concerned, DS2 is a SAS specific language and ASTORE, also proprietary, is used to support algorithms that are trained offline but which are scored in-stream. Model Manager, which is a separate product, supports PMML (predictive modelling mark-up language) and will convert supported model types (more than a dozen currently) into SAS code for deployment on the ESP Server. There is also support for RESTful APIs to run other models. Python notebooks are supported to drive the ESP engine, publish events to ESP, and display results, and models written in Python and C are also supported, though not Java-based models. R will be supported later during 2018. Additional facilities in the ESP Server that are worth mentioning include in-stream geofencing, text analytics, in-stream time pattern recognition (including time-series similarity analysis and time-series clustering), and the ability to build data quality rules into the streaming process.

Of the other elements within the ESP environment, ESP Studio provides an environment for constructing visual models, designed for use by non-technical personnel, while Streamviewer provides a visual analytic dashboard environment that lets you combine real-time and historic data. Event Stream Manager is used to update or deploy analytic algorithms without requiring any downtime on the server, and similarly, add new ESP servers as required. It will automatically discover new servers using ESP agent technology for automated monitoring.

SAS provides connectivity to over 300 end points and supports a variety of standard protocols including MQTT, BACNet publisher connector and adaptor (for smart homes), OPC-UA connector and adaptor (for machine-to-machine communications), a UVC connector (Video4Linux), a WebSocket publisher connector, and a URL publisher connector (for RSS news feeds, JSON from a weather service or News from an HTML page). There are also facilities provided so that you can write your own connectors. In this context it is worth mentioning the new SAS ESP Community (communities.sas.com/IOT), which is moderated by SAS but user-driven.

Machine learning and the Internet of Things are top of mind for many chief data officers and SAS has significant expertise in both of these areas. It is the world’s largest business intelligence and analytics company. It should therefore not come as any great surprise that the company is exceptionally strong when it comes to the breadth of analytic capabilities it provides, regardless of whether this is for data at rest or data in motion. Put this together with the fact that Mckinsey has predicted that the Internet of Things will generate $11 trillion of economic impact by 2025 and you have significant opportunities for SAS to exploit its strength in analytics, across a range of IoT-based deployments. The company is also actively researching analytics for Blockchain, another environment which is set to have a major impact.

The Bottom Line

As a company, SAS is long-established and has a reputation for enterprise-level capability. At the same time it has been a leading light in the analytics market for 40 years. In our view, that makes SAS a major contender, if not a leader, in the market for streaming analytics platforms.

Perhaps the biggest differentiator for SAS is its analytic capabilities, though it’s also notable for its performance. A key feature is the continuous improvement of in-stream models using machine learning. To this end, ESP Viya (though not ESP Lightweight) integrates with SAS Model Manager, a separate product that supports PMML (Predictive Modelling Mark-up Language) and will convert supported model types into SAS code for deployment on ESP. There is also support for RESTful APIs to run other models. Python notebooks can be used to drive the ESP engine, publish events to ESP, and display results, and models written in Python, R, C and Java are also supported. In addition, the product supports ONNX models, which in turn permits the use of PyTorch, Tensorflow, and so on, as well as providing native support for (open source) deep learning. Universal inferencing is also available.

Additional facilities that are worth mentioning include in-stream geofencing, text analytics, in-stream time pattern recognition (including time-series similarity analysis and time-series clustering), and the ability to build data quality rules into the streaming process. Both event-based and window-based (time sliced) processing options are available, and the product boasts full lifecycle support, including health monitoring and high availability. The platform operates in-memory, and a built-in event load manager enables optimised and distributed processing.

Fig 2 - SAS ESP Studio

Of the other elements within the ESP platform, ESP Studio (see Figure 2) provides an environment for constructing visual models, designed for use by non-technical personnel, while Streamviewer provides a visual analytic dashboard environment that lets you combine real-time and historic data. Event Stream Manager is used to update or deploy analytic algorithms without requiring any downtime on the server, and similarly, to add new ESP servers as required. Moreover, ESP Studio, Streamviewer and Event Stream Manager are all integrated into the ESP server Kubernetes environment, thereby supporting ESP deployment, monitoring and updating in that context.

Via built-in ESP Connectors, SAS provides connectivity to over 300 end points, and supports a variety of standard protocols including MQTT, BACNet publisher connector and adaptor (for smart homes), OPC-UA connector and adaptor (for machine-to-machine communications), a UVC connector (Video4Linux), a WebSocket publisher connector, and a URL publisher connector (for RSS news feeds, JSON from a weather service or News from an HTML page). There are also facilities provided so that you can write your own connectors. In this context it is worth mentioning the SAS ESP Community (communities.sas.com/IOT), which is moderated by SAS but user-driven.

SAS is the world’s largest business intelligence and analytics company. It should therefore not come as any great surprise that the company is exceptionally strong when it comes to the breadth of analytic capabilities it provides, regardless of whether it is for data at rest or data in motion. Simply put, SAS has very highly regarded analytics capabilities in general, and via ESP those capabilities can be applied to streaming data just as well as historic data. The company’s expertise when it comes to machine learning and the Internet of Things is a significant draw as well, considering their popularity within, and relevance to, the streaming analytics space.

The Bottom Line

We’ve said it before, and doubtless we’ll say it again: SAS has a reputation for providing enterprise-level quality, and is long-established as a leading light in the analytics market. In our view, this makes the company a major contender within streaming analytics.

Model Manager allows you to store, manage and generally govern all of your models in a single, central location. Emphasis on ‘all’: you do not have to develop your models in SAS to register them in Model Manager. Notably, this means that Model Manager is compatible with models created using open source tools such as R and Python. PMML (predictive modelling mark-up language) is also supported. The product includes model versioning and version control, as well as history tracking. Models can be grouped into projects, and both individual models and projects can be shared between users or team members to enable collaborative working.

Model Manager is accessible via a web browser, through which you can register new models and manage existing ones. ‘Manage’, in this case, includes monitoring (to ensure that models are performing as well as you would like), reporting, deployment, and retraining, among other things. To support these functions, a dashboard is provided to display a variety of analytics and metrics relating to your models, and how they are performing.

On the subject of (re)training, Model Studio, a secondary product available from SAS, allows you to build pipelines for training your models. When you send a model to be retrained, it is reinserted into the appropriate pipeline with new training data attached (alternatively, you can simply flag poorly performing models for a data scientist to look at, and leave it at that). As alluded to above, this can be done at will and ad-hoc.

Model Manager is also designed to promote model reuse. This is the idea that any given model, or variants thereof, should only be developed once, but can (and often should) be deployed multiple times to multiple locations. Accordingly, models in Model Manager can a) be deployed as many times as you want and b) can be deployed to a variety of targets, including databases, data lakes, and streams (including SAS Event Stream Processing), as well as within the SAS platform itself. Moreover, models can be deployed to multiple targets at once with a single click.

Fig 02 - Decision flows and workflows in SAS Model Manager

Having stored and governed your models within Model Manager, SAS provides two products for integrating those models within your business processes. The first of these is SAS Decision Manager, a framework for creating deployable decision flows that incorporate analytics models, policy rules and business logic. The second is SAS Workflow Manager, which allows you to create bespoke, automated workflows that include analytics models. Both decision flows and workflows, which are illustrated in Figure 2, are created using drag and drop interfaces within their respective products, and the latter in particular includes notifications and tasks to prompt expedient workflow completion.

SAS also provides functionality for creating and managing tests for your models, as well as model output validation and scoring. In particular, this includes ‘publishing validation’: testing (and therefore validating) your model as it runs on a particular environment with a particular data source. In other words, you can test in the exact conditions your model is going to be running in. This has obvious advantages.

Finally, SAS supplies a variety of visual analytics which are applicable to models and help with explainability. Most notably, this includes a decision tree diagram, as well a ‘root cause’ analysis of your model’s structure.

Analytics, however sophisticated, does not provide business value in and of itself. It must be put into practice in some way – whether via self-service, building it into workflows, or whatever – to influence and thereby enhance your business processes and decision making. This is just as true for (predictive) models as it is for any other kind of business intelligence. Hence, SAS’ model management offering is valuable for its ability to take the models you have already created and rapidly deploy them such that they are deeply integrated with your existing processes. This enables you to maximise the value you are getting out of the models that you have built.

Moreover, SAS provides a great deal of governance for your models. Although this is not necessarily glamorous, the fact remains that AI and machine learning are only becoming more popular. Consequently, analytical models are becoming more and more widespread and abundant: in due course, we expect large enterprises to be deploying thousands if not tens of thousands of models. What’s more, models can become out-of-date in a matter of months and need replacing or updating on a regular basis. There is simply no way to keep track of – let alone govern, deploy, and maintain – such a proliferation of models without proper tooling, such as the model management products provided by SAS.

The Bottom Line

SAS’ model management offering provides two essential functions for the analytics lifecycle: it manages and governs your models centrally, and it provides the means to operationalise them efficiently. SAS also has one of the most mature solutions in the space, and it shows.