Rollingspanlatency Processor
contrib
Maintainers: @atoulme, @rnterbush, @pavankrish123
Source: opentelemetry-collector-contrib
Supported Telemetry
Overview
Labels spans asslow or very_slow based on a per-key rolling latency baseline, so that
downstream tail-sampling, alerting, or routing decisions can react to spans that are anomalously
slow relative to their own historical behavior rather than a fixed static threshold.
For each incoming span, the processor computes a key from the span name and a configurable set of
resource attributes (e.g. service name, namespace, deployment environment), and maintains an
exponentially-weighted moving average (EWMA) of that key’s duration. Once a key has accumulated
enough observations to warm up, spans whose duration deviates from the rolling mean by more than a
configured number of standard deviations are labeled accordingly.
Configuration
Filtering which spans are evaluated
By default, with bothspan_kinds and always_include_root_spans left unset, every span —
including root spans — is tracked and eligible for labeling; neither option excludes anything on
its own. Setting either option switches the processor into filtering mode, where a span is tracked
only if it matches one of the two criteria below:
Spans that don’t match are passed through unmodified and never contribute to a baseline. Note the
second row: setting
always_include_root_spans: true by itself does not add root spans on top of
“track everything” — since span_kinds contributes no matches when it’s empty, the processor
narrows to root spans only. Set span_kinds too if you want root spans plus specific kinds.
Warnings
resource_key_attributes combined with the span name determines the cardinality of tracked
baselines. Choosing attributes with many unique values (e.g. a per-request ID) can lead to
unbounded memory growth; use max_baselines to bound this and watch for the high-cardinality
warning gated by churn_warning_ratio.
Last generated: 2026-11-09