Iter8-kfserving ships with five out-of-the-box metrics. You can extend the set of available metrics by defining custom metrics.
Metrics shipped out-of-the-box with iter8-kfserving are described in the following table.
- Each metric is defined at a per-version level. For example, the
request-count
metric measures the number of requests to a model version; themean-latency
metric measures the mean latency of a model version. - Metrics can be of type
counter
orgauge
. The value of a counter metric never decreases over time. The two metric types are inspired by Prometheus metric types.
Name | Description | Type | Units |
---|---|---|---|
request-count | Number of requests | counter | |
mean-latency | Mean latency | gauge | milliseconds |
95th-percentile-tail-latency | 95th percentile tail latency | gauge | milliseconds |
error-count | Number of error responses | counter | |
error-rate | Fraction of requests with error responses | gauge |
An overview of iter8 metric resource objects is here. Documentation on how to use metrics within experiments is here. Fields in an iter8 metric resource object spec
are documented here.