The next case study is MJTelco. This fictional company has a huge fiber-optic optical network. One thing about this kind of network, is that it's switched not routed. The ability to route light waves on the fly is currently experimental. So, that means MJTelco needs to figure out where data needs to go, and then configure its network, send the data, and reconfigure the network again for the next dataflow. Predicting flows is a key aspect of making a network like this efficient. MJTelco is already using machine learning for some of this work. More and better models is a high leverage activity that would increase their utilization of their fiber, lower cost, and improve the services that they can offer their customers. In this case, machine learning means both process innovation, increasing efficiency, and product innovation, improving their features. Key business points; MJTelco is an optical networking business. They have inexpensive optical hardware, and they are currently focused on undeserved markets. They've already run a project as a proof of concept, so they know that a cloud solution is viable. Their core values include Kaizen also known as continuous improvement, and they have a strategy of over deploying in a region to overcome regional politics. In other words, if there are two groups fighting for access to the optical network, they want to be able to surprise the potential customers by being able to say yes to both. Their immediate business goals include a data-driven infrastructure for real-time analytics, and machine learning. They want to use machine learning to target customers and also to predict flows, so they can preconfigured their network, and make more efficient use of it. Security is important, scalability, and improving their machine learning applications. Here's one solution concept. Optical networks generally can't route data the way electronic networks can. The optical components necessary to switch data are only recently being developed and are still experimental. Consequently, optical networks are formed of point-to-point links. Data's forwarded either by being optically split and enhanced using an erbium based optical amplifier, or by being converted to electrical signals and regenerated to optical signals for the next hop in the link. Converting a signal from light to electrons and back is a slow process relative to optical speeds. So, understanding the delay involved and when the data is needed is important to deciding how to transmit the dataflows. In a full-mesh optical network, any dataflow could be sent directly from one point to another. However, for capacity management, it might make sense to aggregate data from point A, and send it to point B, and then split or regenerate the signals for forwarded data from B to C, rather than sending data separately from A to B, and from A to C. There's a lot of overhead involved in scaling up a network like this. Reconfiguring the topology of such a network in advance of the plan dataflow, is key to managing such a design. Machine learning is used to predict optimal flows for the network. As the network grows and utilization, the ML model should get better, making the network more efficient and that'll keep the IT overhead down. Again, one tip is to separate business from technical requirements. Look for key words and phrases that map to or imply data engineering solutions. Security, reliability, and scalability are key themes along with improving machine learning. There are clues about a solution in their executive statement. For example, we already have a cost and reliability advantage, we need help meeting capacity, scale, and data security goals. So, that's the scale and security message. We need our analyst focused on product improvements and machine learning, not solving operational problems and our data pipeline. So, that speaks to IT overhead goals. This is interesting, scaling and security, data analyst innovate for customers, so we need rapid and iterative dev test. They need separate development staging and production environments for big data and machine learning. That seems to be a key to this design. Separating these environments, will give them security, speed, and allow each part to grow and scale independently. Rapid scaling. They want to store and analyze two years of data on a continuous basis with a planned rate of 100 million records per day. Security, transport, and storage of data. Some of these technical requirements are typical for machine learning. They're getting useful results, but the time involved affects the value. I recall a story about a customer that was using machine learning to provide extra services to its hotel guests. The only problem was the original algorithm took over a week to run, that meant most of the customers were already checked out when the advice came in to help them. Similarly, the optical network can improve its performance by making better decisions, but if the advice comes after the customer data already needs to arrive, it isn't useful. Very often new companies leveraging machine learning come to realize that timely results are critical and worth it. This is where Cloud ML Engine can be a big help. This case does not give a lot of details about the existing environment or the application itself. All we know is that it's a machine-learning application that relies on big data processing, and that it has to scale, be secure, and involves ML analytics. Therefore, your exam tip in this circumstance is to confine your solution to those elements where you do have information and make reasonable assumptions about the rest. Here are the technical watchpoints; in this case, there are a lot more questions than answers. There's a desire for modular solution that can be iterated when there are new markets, new locations, new clients. Now, it's time for you to define your solution. You might want to read the case study again now that we've highlighted important elements. After you define your solution, you can compare it to a sample solution. Again, the solution that follows is just an example, it's not authoritative. Your solution might be better.
