Foreword
Monitoring is an important part of the guarantee system stability, in Kubernetes open source ecology, resource monitoring tools and components of flourishing. In addition to the communities themselves hatched metrics-server, to the CNCFgraduation Prometheus, developers have a lot of alternative solutions. However, only monitoring resource class is not enough, because of the following two major shortcomings resource monitoring:
- Timeliness and accuracy of monitoring inadequate
Most are based resource monitoring push or pull data off-line mode, it is often a data acquisition time to time, if some burrs or anomalies within a time interval, and resume at the next acquisition is reached, most of the acquisition system will swallow the exception. And for burr scene, the stage of collection will automatically clipping, resulting in reduced accuracy.
- Inadequate monitoring of the scene coverage
The scene is part of the monitoring resources can not be expressed, such as start-stop Pod, we can not simply use resource utilization to measure, because when the resource reaches zero, we can not distinguish the real reason this is not generated by the state.
base