Hello Everyone, I wanted to reach out to the community to discuss potentially enhancing NiFi to include predictive analytics that can help users assess and predict NiFi behavior and performance. Currently NiFi has lots of metrics available for areas including jvm and flow component usage (via component status) as well as provenance data which NiFi makes available either through the UI or reporting tasks (for consumption by other systems). Past discussions in the community cite users shipping this data to applications such as Prometheus, ELK stacks, or Ambari metrics for further analysis in order to capture/review performance issues, detect anomalies, and send alerts or notifications. These systems are efficient in capturing and helping to analyze these metrics however it requires customization work and knowledge of NiFi operations to provide meaningful analytics within a flow context.
In speaking with Matt Burgess and Andy Christianson on this topic we feel that there is an opportunity to introduce an analytics framework that could provide users reasonable predictions on key performance indicators for flows, such as back pressure and flow rate, to help administrators improve operational management of NiFi clusters. This framework could offer several key features: - Provide a flexible internal analytics engine and model api which supports the addition of or enhancement to onboard models - Support integration of remote or cloud based ML models - Support both traditional and online (incremental) learning methods - Provide support for model caching (perhaps later inclusion into a model repository or registry) - UI enhancements to display prediction information either in existing summary data, new data visualizations, or directly within the flow/canvas (where applicable) For an initial target we thought that back pressure prediction would be a good starting point for this initiative, given that back pressure detection is a key indicator of flow performance and many of the metrics currently available would provide enough data points to create a reasonable performing model. We have some ideas on how this could be achieved however we wanted to discuss this more with the community to get thoughts about tackling this work, especially if there are specific use cases or other factors that should be considered. Looking forward to everyone's thoughts and input. Thanks, -yolanda -- yolanda.m.da...@gmail.com @YolandaMDavis
