otelcol-config.yaml

receivers:
  otlp:
    protocols:
      http:
        # The default is localhost:4318
        # But in docker compose, that does not bind to the docker network interface.
        # So we have to use 0.0.0.0 here.
        endpoint: 0.0.0.0:4318

processors:
  attributes:
    actions:
      # Set deployment.environment.name so applications do not have to set it themselves with OTEL_RESOURCE_ATTRIBUTES.
      - key: deployment.environment.name
        value: local
        action: upsert

  # Stamp which otelcol replica processes this data, using OTEL_K8S_POD_NAME,
  # which the opentelemetry-collector Helm chart injects by default via the
  # downward API (metadata.name). This is only STABLE across restarts if the
  # chart is deployed with `mode: statefulset` (ordinals like otel-collector-0)
  # rather than `deployment` (random suffix per restart). In local dev
  # (docker-compose, no Kubernetes), we set it to a fixed value ourselves,
  # see docker-compose.yaml.
  # This runs before transform/metrics_resource_allowlist and
  # deltatocumulative below, so that every metric this replica processes
  # carries its own ID. That makes each replica's deltatocumulative
  # accumulator self-contained: a given (metric, project_id,
  # collector.replica_id) combination can only ever be produced by this one
  # replica, so no consistent/sticky routing is required upstream (plain
  # round-robin across otelcol replicas is fine).
  # Query with sum by (project_id) (increase(...)) to add the replicas'
  # partial series back together.
  resource/collector_replica_id:
    attributes:
      - key: collector.replica_id
        value: ${env:OTEL_K8S_POD_NAME}
        action: insert

  # Allowlist resource attributes for metrics.
  # host.name, container.id, service.version etc. churn on every pod
  # recreation/deploy. With delta temporality, that churn creates new
  # series (and new deltatocumulative accumulator state in this collector)
  # every time, so we only keep service.name (and collector.replica_id,
  # added above) here.
  # deployment.environment.name does not need to be listed: it is set by
  # the `attributes` processor above at the metric-datapoint level, not on
  # the Resource, so it is unaffected by this resource-scoped allowlist.
  transform/metrics_resource_allowlist:
    metric_statements:
      - context: resource
        statements:
          - keep_keys(attributes, ["service.name", "collector.replica_id"])

  # Prometheus (and prometheusremotewrite) only understands cumulative
  # counters, so convert the delta samples we now emit back to cumulative
  # here. Must run after transform/metrics_resource_allowlist so that
  # series from different pods of the same service (but the same
  # collector.replica_id) collapse into one accumulator instead of one per pod.
  deltatocumulative:
    max_stale: 15m

exporters:
  prometheusremotewrite:
    endpoint: http://prometheus:9090/api/v1/write
    tls:
      insecure: true
    resource_to_telemetry_conversion:
      enabled: true
  otlp/tempo:
    endpoint: tempo:4317
    tls:
      insecure: true

service:
  pipelines:
    metrics:
      receivers: [otlp]
      processors: [attributes, resource/collector_replica_id, transform/metrics_resource_allowlist, deltatocumulative]
      exporters: [prometheusremotewrite]
    traces:
      receivers: [otlp]
      processors: [attributes]
      exporters: [otlp/tempo]