metric auto-detection

[lxning]

🚀 The feature

Backend emits metrics via backend metrics APIs. Frontend is able to detect a metrics which is not defined in metrics.yaml and add this metrics in frontend metric cache. This feature workflow is

detect a new metrics name => extract dimension name, unit, type => add this metric definition in frontend cache.

Motivation, pitch

Metric refactor phase2 - support backend model metric auto-detection

In phase 1, TorchServe supports static metrics (ie. metrics defined in metric.yaml), publishes the metrics in either log format or prometheus format.

In phase 2, TorchServe will support the backend metrics not defined in metric.yaml. In this case, it causes perf overhead in frontend.

Alternatives

No response

Additional context

No response