Connecting OpenTelemetry traces to Prometheus

Operation traces provide insight into performance issues that are occurring at various execution points in your graph. However, individual traces don't provide a view of your graph's broader performance.

Helpfully, you can convert your operation traces into aggregated metrics without requiring manual instrumentation. To accomplish this, we'll use spanmetricsprocessor in an OpenTelemetry Collector instance to automatically generate metrics from our existing trace spans.

OpenTelemetry Collector configuration

OpenTelemetry provides two different repositories for their OpenTelemetry Collector:

• The
  core library
• The
  contributor library

These repositories are similar in scope, but the contributor library includes extended features that aren't suitable for the core library. To derive performance metrics from our existing spans, we'll use the contributor library to take advantage of the spanmetricsprocessor via the associated Docker image.

When your OpenTelemetry Collector is ready to run, you can start configuring it with this barebones example:

1
receivers:
2
  otlp:
3
    protocols:
4
      grpc:
5
      http:
6
        cors:
7
          allowed_origins:
8
            - http://*
9
            - https://*
10
  otlp/spanmetrics:
11
    protocols:
12
      grpc:
13
        endpoint: 0.0.0.0:12346
14

15
exporters:
16
  prometheus:
17
    endpoint: '0.0.0.0:9464'
18

19
processors:
20
  batch:
21
  spanmetrics:
22
    metrics_exporter: prometheus
23

24
service:
25
  pipelines:
26
    traces:
27
      receivers: [otlp]
28
      processors: [spanmetrics, batch]
29
    metrics:
30
      receivers: [otlp/spanmetrics]
31
      exporters: [prometheus]
32
      processors: [batch]

Apollo Server setup

Add the OTLP Exporter (@opentelemetry/exporter-trace-otlp-http Node package) following the same instructions

Apollo Router setup

To send traces from the Apollo Router to OpenTelemetry Collector, see

Prometheus setup

Lastly, we need to add the OpenTelemetry Collector as a target within Prometheus. It'll use the standard port for Prometheus metrics (9464).

That's it- you should have access to span metrics using the same operation name!

Example queries

Here are a few sample queries to help explore the data structure being reported:

• P95 by service: histogram_quantile(.95, sum(rate(latency_bucket[5m])) by (le, service_name))
• Average latency by service and operation (for example router / graphql.validate): sum by (operation, service_name)(rate(latency_sum{}[1m])) / sum by (operation, service_name)(rate(latency_count{}[1m]))
• RPM by service: sum(rate(calls_total{operation="HTTP POST"}[1m])) by (service_name)

Full demo

To see this in action, check out the