Ingestion volume control with APM Distributed Tracing

Overview

The Ingestion control page provides granular visibility into the ingestion configuration for all services, in the agent and in the tracing libraries. All Ingestion Mechanisms are publicly documented and configurable.

With the ingestion control page, you have full visibility and complete control of your span volume. Consequently, you are be able to:

  • Ingest the data that is most relevant to your business and your observability goals.
  • Reduce network costs by avoiding sending unused trace data to the Datadog platform.
  • Control and manage your overall costs.

Effects of reducing trace ingestion volume

APM ingestion sampling displaying 25 percent complete traces ingested

If you decide to reduce the ingestion volume for certain services, the request, error, and latency metrics (known as RED metrics, for Requests, Errors, and Duration) remain 100% accurate, as they are being calculated based on 100% of the application’s traffic, regardless of any sampling configuration. These metrics are included when purchasing Datadog APM. In order to make sure you have full visibility into your application’s traffic, you can use these metrics to spot potential errors on a service or a resource, by creating dashboards, monitors, and SLOs.

Note: If your applications and services are instrumented with OpenTelemetry libraries and you set up sampling at the SDK level and/or at the collector level, APM metrics are based on the sampled set of data.

Trace data is very repetitive, which means trace samples to investigate any issues are still available with ingestion sampling. For high throughput services, there’s usually no need for you to collect every single request - an important enough problem should always show symptoms in multiple traces. Ingestion controls helps you to have the visibility that you need to troubleshoot problems while remaining within budget.

Metrics from spans

Metrics from spans are based on ingested spans.

Reducing ingestion sampling rates will impact any count type metric. Distribution type metrics, for instance duration measures, are not impacted as the sampling is mostly uniform, the distribution of latencies remains representative of the traffic.

Monitors

Any metric monitor using metrics from spans is impacted by ingestion volume reduction. Metric monitors based on trace.__ metrics will remain accurate, because these metrics are calculated based on 100% of the traffic.

Count-based Trace analytics monitors are impacted as well. Check if you have trace analytics monitors created by looking for type:trace-analytics monitors in the manage monitors page.

Assess your services’ ingestion configuration

To assess the current state of applications’ instrumentation, leverage the Trace Ingestion Control page that provides detailed information on agent and tracing library configuration.

Understanding if you are within your monthly ingestion allocation

Use the ingestion monthly usage KPI to get an estimation of your usage compared to the monthly allocation of 150 GB of ingested spans per APM host (summed across all APM hosts).

Ingestion Overage KPI displaying 170 percent estimated monthly usage of 23.3 monthly available TB across all infrastructure

Advanced APM usage investigation

The ingestion configuration can be investigated for each service. Click on a service row to see the Service Ingestion Summary, which surfaces:

  • Ingestion reason breakdown: which ingestion mechanism is responsible for the ingestion volume
  • Top sampling decision makers: which upstream services are taking sampling decisions for the spans ingested in regards to the default ingestion mechanism

An out-of-the-box dashboard is also available to get more insights on historical trends related to your ingestion usage and volume. Clone this dashboard to be able to edit widgets and perform further analysis.

Reduce your ingestion volume

Identify services responsible for most of the ingestion volume

To identify which services are responsible for most of the ingestion volume, sort the table by Downstream Bytes/s. This column allows you to spot which services take most of the sampling decisions, which also impact downstream services.

If the service is starting the trace, Downstream Bytes/s also encompasses the volume of spans coming from downstream services for which the service took the sampling decision.

The Traffic Breakdown column gives a good indication of the service’s sampling configuration.

If the service has a high Downstream Bytes/s rate and a high sampling rate (displayed as the blue filled section of the traffic breakdown column), reducing the sampling rate for this service is expected to have a high impact on the ingestion volume.

APM ingestion sampling displaying 99 percent complete traces ingested, meaning no sampling

Globally configure the ingestion sampling rate at the Agent level

The Configuration column tells you whether or not your services are configured with sampling rules. If the top services are labelled with AUTOMATIC configuration, changing the Agent configuration will reduce the volume globally accross services.

To reduce the ingestion volume at the Agent level, configure DD_APM_MAX_TPS (set to 10 by default) to reduce the share of head-based sampling volume. Read more about the default sampling mechanism.

Note: This configuration option only goes into effect when using Datadog tracing libraries. If the OTLP Ingest in the Agent collects data from applications instrumented with OpenTelemetry, modifying DD_APM_MAX_TPS does not change sampling rates that are applied in tracing libraries.

Additionally, to reduce the volume of error and rare traces:

  • Configure DD_APM_ERROR_TPS to reduce the share of error sampling.
  • Set DD_APM_DISABLE_RARE_SAMPLER to true to stop sampling rare traces.

Independently configure the ingestion sampling rate for services at the library level

By configuring sampling rates for a few high-throughput services, most of the “exceeding” ingestion volume can be lowered.

Click on a service to view the Service Ingestion Summary. Look at the Ingestion reasons breakdown in the side panel, which gives an overview of the share of ingestion volume attributed to each mechanism.

If the main reason for most of the ingestion volume is head-based sampling (auto or rule), the volume can be configured by setting a sampling rule at the tracing library level.

Click the Manage Ingestion Rate button to configure a sampling rate for the service. Select the service language and the ingestion sampling rate you want to apply.

Note: The application needs to be redeployed in order to apply the configuration changes. Datadog recommends applying the changes by setting environment variables.

Trace sampling with OpenTelemetry

If your applications and services are instrumented with OpenTelemetry libraries and you’re using the OpenTelemetry collector, you can use the following OpenTelemetry sampling capabilities:

Using either of the two options results in sampled APM Metrics.

Ingestion reasons glossary

Know which ingestion mechanisms are responsible for most of the ingestion volume

The default mechanism to sample traces is head-based sampling. The decision whether to sample a trace or not is taken at the beginning of its lifecycle, and propagated downstream in the context of the requests in order to ensure that you can always view and analyze complete traces.

Head-based sampling is configurable in the tracing libraries or from the Datadog Agent:

ingestion reasonWhereIngestion Mechanism DescriptionDefault
autoAgentThe Datadog Agent distributes sampling rates to tracing libraries.10 traces per second per Agent
ruleTracing LibrariesThe libraries’ defined sampling percentage for specific services.null

Several other ingestion reasons are surfaced in the Ingestion Control page and as a tag on the datadog.estimated_usage.apm.ingested_bytes metric. These ingestion reasons may be responsible for your ingestion volume:

ingestion reasonWhereIngestion Mechanism DescriptionDefault
errorAgentSampling of errors uncaught by the head-based sampling.10 traces per second per Agent (null, if rules are defined)
rareAgentSampling of rare traces (catching all combinations of a set of span tags).5 traces per second per Agent (null, if rules are defined)
manualIn-codeIn-code decision override to keep/drop a span and its children.null
analyticsAgent and Tracing LibrariesDeprecated ingestion mechanism that samples single spans without the full trace.null

Additionally, other products can be responsible for sampled span volume:

  • synthetics and synthetics-browser: API and browser tests are connected to the trace generated by the test.
  • rum: Requests from web and mobile applications are linked to the corresponding backend traces.
  • lambda and xray: Traces generated from AWS lambda functions instrumented with X-Ray or Datadog libraries.

Read more about ingestion reasons in the Ingestion Mechanisms documentation.

Further Reading

Additional helpful documentation, links, and articles: