Open
Description
Description
The "ray status" logs periodically emitted by the autoscaler have the following issues:
- (biggest issue) They are emitted once per several seconds, even when the cluster status hasn't changed. This drowns out information about actual events and issues. Only log status when there is a change. The definition of "change" is TBD, but for a first cut, a literal change in the status string (ignoring timestamp) could probably work. This change should be accompanied by an info message and a feature flag to disable it.
- (smaller issue) The logs are multi-line, i.e. they contain newline characters. That's bad for structured logging. We've had some complaints about this. Consider logging the status line-by-line.
Use case
Improved log-viewing experience.