Proposal: Span Stats Processor

[ashwinidulams]

Span Statistics Proposal

A processor for aggregating spans to derive RED (Request, Error, Distribution) metrics across a set of dimensions (dynamic or otherwise) and exporting the results to a metrics sink (prometheus endpoint, stream etc).

Motivation

A trace contains end-to-end data about the service call chain in the context of a request/transaction. This data can be mined to derive context rich metrics that can help us identify and isolate the problem as quickly as possible. For example, span level aggregation can be used to detect any performance problems within the process boundary defined by the span.

Instead of forwarding the spans to a separate processor (stream processing application) for aggregation, a otel processor can process the spans in flight and emit RED metrics at configurable time intervals.

Proposed Configuration

processors:
  # name of the processor 
  spanstatsprocessor:
    # time window for aggregation in seconds.
    timewindowseconds: 15
    # dimensions for aggregation
    dimensions:
    - http.url
    - custom_tag1
    # sink for metrics
    sink:
      # prometheus sink with a endpoint exposed at port 24880
      prometheusport: 24880

• I believe this processor should not be part of the core and should be in contrib repo

[bogdandrutu]

Would suggest this to be a contrib feature. There are a lot of unknowns in this feature so probably better to start there, and once stable and proven functionality we can move it to core

[jmacd]

See open-telemetry/opentelemetry-specification#657, which aims to specify conventions for this kind of processor.

[jrcamp]

Closing since it's filed under open-telemetry/opentelemetry-collector-contrib#403 now.