Watson AIOps Metric Manager

Posted by Keith Posner
Author: Ian Manning

Time series metric data is prolific throughout the world.  In my 11 years of working with time series data in IBM, we have built analytics that have analysed data across realms as varied as IT systems, Telco systems, turbines, and oil wells. Once, just for fun, we even built analytics to analyse time series data from a set of bicycle computers measuring speed, distance, pedal cadence, and heart rate. The best part is that all of this analysis has been performed automatically using our own machine learning analytics.

IBM messaging, and increased competition in the AIOps space, have sensitized customers to the need to detect anomalies to help avoid service impacting problems. Anomalies help you understand what has changed, how quickly it is changing, and how to rapidly resolve the issue.  This type of information is vital when outages are costly, and brand reputation is on the line.

Our algorithms are focused on detecting real-world problems, with a focus on providing value from day one. Scalable, understandable, and noise-free, our time series analytics train automatically and self-tune ­– they don’t ask questions that they can more reliably determine on their own.  This means that you can focus on the core business of monitoring, and can leave data science tasks such as tuning, model maintenance, and model management to the algorithms.

As we move forward, we plan to expand our data coverage by creating time series analytics from log files, from events, and even from topological changes.  These analytics can reveal a range of new insights, such as:

• Abnormal behaviour, based on your log file data.
• Event storms and other indications of greater than normal event counts, based on your event data.
• Abnormal service provisioning and unexpected service replication, based on your topological change data.

In the new Watson AIOps product we have built the time series analytics based on years of research and close collaboration with customers. Two principles have guided us from day one:

• The analytics must be configuration-free.
• The analytics must be data agnostic.

The practical implication is that these time series analytics do not require a data scientist, and do not require information about the data being ingested. Instead, the analytics learn features and patterns from the numbers, and from the numbers alone. In this way, if you have proprietary application metrics, Watson AIOps can analyse them.  If you have metrics from other vendors, Watson AIOps can analyse them.  And the best part is that Watson AIOps can analyse all of the metrics together, at the same time as it monitors all of your other data – logs, events, tickets, change requests, and topology data.