Telemetry and Monitoring

One of Linkerd's most powerful features is its extensive set of tooling around observability—the measuring and reporting of observed behavior in meshed applications. While Linkerd doesn't have insight directly into the internals of service code, it has tremendous insight into the external behavior of service code.

Linkerd's telemetry and monitoring features function automatically, without requiring any work on the part of the developer. These features include:

  • Recording of top-line (“golden”) metrics (request volume, success rate, and latency distributions) for HTTP, HTTP/2, and gRPC traffic.
  • Recording of TCP-level metrics (bytes in/out, etc) for other TCP traffic.
  • Reporting metrics per service, per caller/callee pair, or per route/path (with Service Profiles).
  • Generating topology graphs that display the runtime relationship between services.
  • Live, on-demand request sampling.

This data can be consumed in several ways:

Golden metrics

Success Rate

This is the percentage of successful requests during a time window (1 minute by default).

In the output of the command linkerd routes --o wide, this metric is split into EFFECTIVE_SUCCESS and ACTUAL_SUCCESS. For routes configured with retries, the former calculates the percentage of success after retries (as perceived by the client-side), and the latter before retries (which can expose potential problems with the service).

Traffic (Requests Per Second)

This gives an overview of how much demand is placed on the service/route. As with success rates, linkerd routes --o wide splits this metric into EFFECTIVE_RPS and ACTUAL_RPS, corresponding to rates after and before retries respectively.

Latencies

Times taken to service requests per service/route are split into 50th, 95th and 99th percentiles. Lower percentiles give you an overview of the average performance of the system, while tail percentiles help catch outlier behavior.

Lifespan of Linkerd metrics

Linkerd is not designed as a long-term historical metrics store. While Linkerd's control plane does include a Prometheus instance, this instance expires metrics at a short, fixed interval (currently 6 hours).

Rather, Linkerd is designed to supplement your existing metrics store. If Linkerd's metrics are valuable, you should export them into your existing historical metrics store.

See Exporting Metrics for more.