Skip to main content
Copy page
Version: Cloud

Configure Trace Sampling

At high request rates, exporting every trace becomes impractical — a single agent invocation produces a root span plus one span per LLM call, tool call, and retrieval step.

AMP does not sample traces on the server side. Sampling is a client-side, head-based decision made inside the agent process by the OpenTelemetry SDK, before spans are exported. You control it with two standard OpenTelemetry environment variables:

VariablePurpose
OTEL_TRACES_SAMPLERWhich sampling strategy to use
OTEL_TRACES_SAMPLER_ARGThe sampling rate, for the ratio-based samplers

No code change is required — the OpenTelemetry SDK that the Traceloop SDK builds on reads them at startup.

Python agents

This page covers Python agents instrumented with amp-instrumentation. An agent that configures a sampler in code overrides these environment variables — the .NET setup in AMP Instrumentation does exactly that with .SetSampler(new AlwaysOnSampler()), so change the sampler in code there instead.

Supported Values​

OTEL_TRACES_SAMPLERBehaviour
parentbased_always_onDefault. Keep every trace.
parentbased_traceidratioKeep a fraction of traces, set by OTEL_TRACES_SAMPLER_ARG, but reuse the caller's decision when there is one. Recommended.
traceidratioKeep a fraction of traces, ignoring the caller's decision.
always_onKeep every trace, even ones the caller decided to drop.
always_offDrop everything. Disables tracing for the agent.
parentbased_always_offDrop everything except traces the caller already decided to keep.

OTEL_TRACES_SAMPLER_ARG is a decimal between 0.0 and 1.0 and applies only to the two ratio samplers — 0.1 keeps roughly 10% of traces. Values outside the list above (for example jaeger_remote or xray) are not supported by the bundled SDK.

export OTEL_TRACES_SAMPLER="parentbased_traceidratio"
export OTEL_TRACES_SAMPLER_ARG="0.1"

This keeps about 10% of traces.

Sampling is per trace, not per span. The decision is derived from the trace ID, so every span in a trace shares the same outcome — a sampled trace arrives complete in the Console, and an unsampled one never leaves the agent process. You will not see partial traces.

What "parent-based" means​

When another instrumented service calls your agent, that service starts the trace and decides whether to keep it. It passes that decision along in the standard traceparent HTTP header. The parent is simply the calling span.

  • A parent-based sampler reuses the caller's decision when a caller is present, and applies its own rule only when the agent starts the trace itself.
  • A non-parent-based sampler (traceidratio, always_on, always_off) ignores the caller entirely and always applies its own rule.

Prefer parentbased_traceidratio because it keeps distributed traces whole. If the calling service decided to keep a trace but your agent independently dropped it, the trace would still appear upstream with the agent's portion missing.

If nothing calls your agent through an instrumented service — that is, the agent always starts its own traces — then parentbased_traceidratio and traceidratio behave identically.

Where to Set the Variables​

The variables are the same for both deployment models — only how you set them differs:

Platform-hosted agentExternally-hosted agent
WhereConsole → Deploy → Configurations and Secrets → ConfigureThe agent's own environment
Endpoint and API keyInjected by AMP — you do not set themYou set AMP_OTEL_ENDPOINT and AMP_AGENT_API_KEY yourself
You addThe two sampler variables onlyThe two sampler variables, alongside the endpoint and API key
To applyRedeploy the agentRestart the agent

The sampler is read once when the agent process starts, so a change never affects a running agent until it is redeployed or restarted.

Platform-Hosted Agents​

AMP already injects AMP_OTEL_ENDPOINT and AMP_AGENT_API_KEY, so you only add the two sampler variables:

  1. Open the agent and click Deploy.
  2. On the card for the environment you want to sample, click Configure next to Configurations and Secrets.
  3. In the drawer, scroll to Environment Variables and add OTEL_TRACES_SAMPLER and OTEL_TRACES_SAMPLER_ARG.
  4. Save and redeploy the agent.

You can also set the same variables when first deploying the agent, from the Environment Variables section of the Configure & Deploy step.

Environment variables apply to one environment at a time, so set them for each environment you want sampled; values for later environments are configured when you promote. See Manage Environments and Deployment Pipelines for how promotion works.

This works whether the agent uses auto-instrumentation or the manual init_otel() helper — both read the variables from the container environment at startup.

Externally-Hosted Agents​

Set all four variables in the agent's own environment, then start it:

export AMP_OTEL_ENDPOINT="<your-amp-otel-endpoint>"
export AMP_AGENT_API_KEY="<your-generated-api-key>"
export OTEL_TRACES_SAMPLER="parentbased_traceidratio"
export OTEL_TRACES_SAMPLER_ARG="0.1"

amp-instrument python my_agent.py

The same variables apply if you call init_otel() from your own code instead of using amp-instrument.

Important Considerations​

Invalid values fail open, silently. A misspelled sampler name falls back to the default parentbased_always_on, and a non-numeric OTEL_TRACES_SAMPLER_ARG falls back to a rate of 1.0. Both export every trace — the opposite of what you intended, with only a warning in the agent logs. After changing these variables, confirm in the Console that trace volume actually dropped.

Sampled-out traces are gone for good. They cannot be searched in the Console and are not available to evaluation monitors. A monitor on a 10%-sampled agent scores only that 10% — factor this into how you read monitor scores.

Head sampling is random, not selective. The decision is made when the trace starts, before the agent has done any work, so the SDK cannot preferentially keep errors or slow requests. A 10% sample keeps roughly 10% of failures too. If you need every error, keep full sampling and filter at query time instead.