Transaction sampling

edit

Distributed tracing can generate a substantial amount of data. More data can mean higher costs and more noise. Sampling aims to lower the amount of data ingested and the effort required to analyze that data — all while still making it easy to find anomalous patterns in your applications, detect outages, track errors, and lower mean time to recovery (MTTR).

Elastic APM supports two types of sampling:

Head-based sampling
edit

In head-based sampling, the sampling decision for each trace is made when the trace is initiated. Each trace has a defined and equal probability of being sampled.

For example, a sampling value of .2 indicates a transaction sample rate of 20%. This means that only 20% of traces will send and retain all of their associated information. The remaining traces will drop contextual information to reduce the transfer and storage size of the trace.

Head-based sampling is quick and easy to set up. Its downside is that it’s entirely random — interesting data might be discarded purely due to chance.

See Configure head-based sampling to get started.

Distributed tracing with head-based sampling

In a distributed trace, the sampling decision is still made when the trace is initiated. Each subsequent service respects the initial service’s sampling decision, regardless of its configured sample rate; the result is a sampling percentage that matches the initiating service.

In this example, Service A initiates four transactions and has sample rate of .5 (50%). The sample rates of Service B and Service C are ignored.

Distributed tracing and head based sampling example one

In this example, Service A initiates four transactions and has a sample rate of 1 (100%). Again, the sample rates of Service B and Service C are ignored.

Distributed tracing and head based sampling example two

OpenTelemetry with head-based sampling

Head-based sampling is implemented directly in the APM agents and SDKs. The sample rate must be propagated between services and the managed intake service in order to produce accurate metrics.

OpenTelemetry offers multiple samplers. However, most samplers do not propagate the sample rate. This results in inaccurate span-based metrics, like APM throughput, latency, and error metrics.

For accurate span-based metrics when using head-based sampling with OpenTelemetry, you must use a [consistent probability sampler](https://opentelemetry.io/docs/specs/otel/trace/tracestate-probability-sampling/). These samplers propagate the sample rate between services and the managed intake service, resulting in accurate metrics.

OpenTelemetry does not offer consistent probability samplers in all languages. OpenTelemetry users should consider using tail-based sampling instead.

+ Refer to the documentation of your favorite OpenTelemetry agent or SDK for more information on the availability of consistent probability samplers.

Tail-based sampling
edit

In tail-based sampling, the sampling decision for each trace is made after the trace has completed. This means all traces will be analyzed against a set of rules, or policies, which will determine the rate at which they are sampled.

Unlike head-based sampling, each trace does not have an equal probability of being sampled. Because slower traces are more interesting than faster ones, tail-based sampling uses weighted random sampling — so traces with a longer root transaction duration are more likely to be sampled than traces with a fast root transaction duration.

A downside of tail-based sampling is that it results in more data being sent from APM agents to the APM Server. The APM Server will therefore use more CPU, memory, and disk than with head-based sampling. However, because the tail-based sampling decision happens in APM Server, there is less data to transfer from APM Server to Elasticsearch. So running APM Server close to your instrumented services can reduce any increase in transfer costs that tail-based sampling brings.

See Configure tail-based sampling to get started.

Distributed tracing with tail-based sampling

With tail-based sampling, all traces are observed and a sampling decision is only made once a trace completes.

In this example, Service A initiates four transactions. If our sample rate is .5 (50%) for traces with a success outcome, and 1 (100%) for traces with a failure outcome, the sampled traces would look something like this:

Distributed tracing and tail based sampling example one

OpenTelemetry with tail-based sampling

Tail-based sampling is implemented entirely in APM Server, and will work with traces sent by either Elastic APM agents or OpenTelemetry SDKs.

Due to OpenTelemetry tail-based sampling limitations when using tailsamplingprocessor, we recommend using APM Server tail-based sampling instead.

Sampled data and visualizations

edit

A sampled trace retains all data associated with it. A non-sampled trace drops all span and transaction data1. Regardless of the sampling decision, all traces retain error data.

Some visualizations in the APM app, like latency, are powered by aggregated transaction and span metrics. Metrics are based on sampled traces and weighted by the inverse sampling rate. For example, if you sample at 5%, each trace is counted as 20. As a result, as the variance of latency increases, or the sampling rate decreases, your level of error will increase.

1 Real User Monitoring (RUM) traces are an exception to this rule. The Kibana apps that utilize RUM data depend on transaction events, so non-sampled RUM traces retain transaction data — only span data is dropped.

Sample rates

edit

What’s the best sampling rate? Unfortunately, there isn’t one. Sampling is dependent on your data, the throughput of your application, data retention policies, and other factors. A sampling rate from .1% to 100% would all be considered normal. You’ll likely decide on a unique sample rate for different scenarios. Here are some examples:

  • Services with considerably more traffic than others might be safe to sample at lower rates
  • Routes that are more important than others might be sampled at higher rates
  • A production service environment might warrant a higher sampling rate than a development environment
  • Failed trace outcomes might be more interesting than successful traces — thus requiring a higher sample rate

Regardless of the above, cost conscious customers are likely to be fine with a lower sample rate.