DDSketch-based Metrics in APM

Trace metrics are collected automatically for your services and resources and are retained for 15 months. The latency percentiles exist as individual timeseries. These percentiles are also available as a Datadog Distribution Metric. Instead of having a different metric for each percentile; and separate metrics for services, resources, or second primary tags; Datadog offers a simple metric:

  • trace.<SPAN_NAME>:
    • Prerequisite: This metric exists for any APM service .
    • Description: Represents latency distributions for all services, resources and versions across different environments and second primary tags.
    • Metric type: DISTRIBUTION
    • Tags: env, service, version, resource, and the second primary tag.

The APM Service and Resource pages use this metric type automatically. You can use these metrics to power your dashboards and monitors.

How am I seeing a full history of this new metric?

  • Datadog stitches any existing query on the new metric to an equivalent query based on the long-existing latency metrics, so you do not have to create multiple queries.

I’m seeing a change in the values of my latency, what is happening?

  • Datadog Distribution Metrics are powered by DDSketch. This includes a change from rank-error guarantees to relative error guarantees. As a result, all percentile estimate values are now guaranteed to be closer to the true percentile value.
  • Specifically, you might expect to see a reduction in p99 values, where this difference is most noticeable. The new values are centered more closely on the precise p99 value.
  • One thing to note is that the APM metric calculations are not exactly akin to a Datadog Distribution Custom Metric that would be calculated in-code. The calculation happens on the backend, so some differences may occur.

I’m using Terraform. What does this change mean for me?

  • The existing metrics are still around; your Terraform definitions are still in place and still work.
  • To take advantage of the better precision offered by the new DDSketch-based metrics, change your Terraform definitions as shown in the following examples.

Percentiles before:{service:foo, resource:abcdef1234}{datacenter:production, service:bar, resource:ghijk5678}

Percentiles after:

p99:trace.http.request{service:foo, resource:abcdef1234}
p75:trace.sample_span{datacenter:production, service:bar, resource:ghijk5678}

p100 before:{service:foo, resource:abcdef1234}{datacenter:production, service:bar, resource:ghijk5678}

p100 after:

max:trace.http.request{service:foo, resource:abcdef1234}
max:trace.sample_span{datacenter:production, service:bar, resource:ghijk5678}