Use case
Jev for Metric retention
Ask Jev whether an unfamiliar metric looks worth keeping, then let Collector policy apply the threshold.
Last verified: 2026-09-30
Overview
A new service can emit metrics nobody has classified. Usage history does not exist yet, and a hand-written allowlist only covers names you already know.
Why Jev fits
- Names, descriptions, units, and attribute keys are enough state for a Noul.
- The keep-or-drop line stays in Collector config, so a probability is not itself a deletion.
- A choice recommendation can be stored for review while only one Noul is allowed to filter.
- Cached answers keep the next batch from paying for the same metadata again.
Real projects
Example workflow
- Protect known-critical metric names before any model call.
- Send metadata for the rest to hosted Jev.
- Start by annotating the stream so you can read the scores.
- Turn on reduction only after the keep threshold matches what you want to delete.
When Jev works well
- Pipelines where the first question is whether a metric is operationally useful.
- Teams that can review assessments before anything is dropped.
When Jev may not fit
- Cardinality and cost decisions that depend on query volume the metric metadata does not contain.
- A pipeline that must drop on the first batch. The published processor queues inference and applies it later.
Related tutorials
NoulComposite scoringjevmetrics
Sources
- jevmetrics READMEishantanu · accessed 2026-09-30 · github
- NoulTypeSafe · accessed 2026-09-30 · documentation