Thanks a lot for the feedback, Sebastian.
As for the data being pushed it would expect that would contain metrics, events and topology information, but that is certainly a topic to be discussed and to which I don't have a clear answer at this time.
Also, I'm well aware that it's not possible to predict every failure - it would be naive to assume that.
I think one of the assumptions here is that there might be significant untapped potential in combining/augmenting APM data with "foreign" domains such as marketing, social and even physical domains. In this case data-agnostic AIOps solutions would act as an aggregator for different domain-centric data. In the end one could perhaps imagine a solution implementing a general business operations support model not unlike that used for weather forecasts 🙂 After all, it's not "only" about predicting application outages but for example also user behavior and market dynamics.
But even if we stick to the APM domain there are scenarios where outages could be predicted with some confidence using very simple algorithms. Think for example of linear timeseries extrapolation to detect possible, future threshold breaches (i.e. increasing resource usage such as disk, memory, object pools etc.). Of course there is a risk for over-alerting unless the AI model has deeper knowledge about the observed system (i.e. code analysis, configuration, common fault patterns etc.) but this is only to illustrate a simple example for a "predictive capability" that would have given us a head start in some cases (admittedly one would not need an AI for that)...
One use case we are pursuing going forward is the automatic merging and correlation of/with arbitrary event, topology and CMDB data streams obtained from different "silos" and domains within our company (Dynatrace being only one of them) and thus enabling true domain-agnostic AIOps.
