Auto-scaling SIG/Theory of Auto-Scaling
Contents
Theory of Auto-Scaling
General Description
<fill in> <what is the scope of auto-scaling, how does it differ from self-healing, what does it have in common with self-healing>
Conceptual Diagram
Components of Auto-Scaling
OpenStack offers a rich set of services to build, manage, orchestrate, and provision a cloud. This gives administrators some choices in how to best serve their customer's needs.
- Scaling units - There are a number of components that can be controlled with Auto-Scaling.
- Compute Host
- VM running on a Compute Host
- Container running on a Compute Host
- Network Attached Storage
- Virtual Network Functions
- Monitoring Service - either using an agent installed on the Scaling unit, or using a polling method to retrieve metrics
- Monasca
- Ceilometer from the Telemetry project
- Prometheus
- Alarming Service
- Monasca has a built in alarm thresholding service and notification service
- Aodh from the Telemetry project
- Decision Services - There are a number of services in OpenStack that can interpret metrics and alarms based on configured logic and produce commands to Orchestration Engines
- Congress
- Heat
- Vitrage
- Watcher
- Orchestration Engines
- Heat
- Senlin is a clustering engine for OpenStack, and can orchestrate auto-scaling
- Tacker
Considerations and Guidelines
- Monitoring takes resources, plan accordingly
- Avoid scaling too quickly or too often
- This can be done by specifying appropriate cooldown periods.
- Another technique is to average the scaling metric over a longer time period to avoid reacting to sudden fluctuations
- Don't expect instantaneous scaling (see above)
- Define thresholds to be predictive of scale needs, not reactive to a bad state
- Be aware of where the logic for scaling is (alarm thresholds, decision services)
- Define appropriate scaling limits in terms of minimum and maximum instances.
- Minimum number of instances will prevent all the instances from being removed.
- Maximum number of instances safeguards against provisioning too many resources that could adversely affect other workloads.
- Applications must be horizontally scalable in order to auto-scale the underlying instances.
- Applications must be stateless or be able to drain existing stateful connections so that the underlying instances can be removed during a scale down.
- Incoming requests must be dynamically load balanced among the instances running the application.