Jevmetrics – probabilistic metric reduction inside OpenTelemetry

Jevmetrics is an experimental OpenTelemetry Collector connector that uses TypeSafe’s Jev probabilistic model to infer how operationally valuable metric instruments are from their metadata, then enforces deterministic retention policies based on those probabilities. It runs inference asynchronously, caches assessments, and supports an annotate mode that preserves incoming metrics while exporting the inferred keep/drop assessments for human review. The connector constructs typed questions from metric name, description, units, instrumentation scope, temporality and monotonicity, datapoint and distinct-series counts, and selected resource values; raw datapoint values, histogram buckets, and exemplars are not sent. Decisions use configurable keep and drop thresholds, with fail-open behavior for unscored metrics and exponential backoff on inference failures. Cache identity includes full resource and instrumentation details; LRU eviction and expiry are used and restart clears the cache.

The project emphasizes placing semantic inference inside the telemetry pipeline to complement explicit filter rules and usage-based systems, enabling early decisions for unfamiliar instrumentation before historical usage exists. Limits are explicit: Jev evaluates single-metric summaries without history, dashboards, alerts, or cross-metric comparison, and no published accuracy, signal-loss, or cost-savings measurements exist yet. Recommended workflow is to run annotate, compare assessments to engineer review, protect critical metrics, and measure overhead before enabling filtering. Implementation is in Go, includes tests, a Collector build, CI with mock responses, and an Apache-2.0 license.