/* Span Creation scenarios. This benchmark measures the performance cost of different span creation patterns in OpenTelemetry when using On/Off sampling strategies. The benchmark results: criterion = "0.5.1" cargo 1.91.1 (ea2d97820 2025-10-10) Hardware: M4Pro | Test | Always Sample | Never Sample | |-------------------------------------------------------|---------------|--------------| | span-creation-simple | 203.67 ns | 51.316 ns | | span-creation-span-builder | 209.55 ns | 88.032 ns | | span-creation-tracer-in-span | 339.52 ns | 150.55 ns | | span-creation-tracer-in-span-with-builder | 317.44 ns | 170.38 ns | | span-creation-simple-context-activation | 329.41 ns | 47.681 ns | | span-creation-span-builder-context-activation | 328.84 ns | 78.021 ns | */ use criterion::{criterion_group, criterion_main, Criterion}; use opentelemetry::{ trace::{mark_span_as_active, Span, SpanBuilder, TraceContextExt, Tracer, TracerProvider}, Context, KeyValue, }; use opentelemetry_sdk::{ error::OTelSdkResult, trace::{self as sdktrace, SpanData, SpanExporter}, }; #[cfg(all(not(target_os = "windows"), feature = "bench_profiling"))] use pprof::criterion::{Output, PProfProfiler}; fn criterion_benchmark(c: &mut Criterion) { trace_benchmark_group(c, "span-creation-simple", |tracer| { // Simple span creation // There is not ability to specify anything other than the name. // Attributes are set after creation, and automatically gets // ignored if sampling is unfavorable. let mut span = tracer.start("span-name"); if span.is_recording() { span.set_attribute(KeyValue::new("key1", false)); span.set_attribute(KeyValue::new("key2", "hello")); span.set_attribute(KeyValue::new("key3", 123.456)); span.set_attribute(KeyValue::new("key4", "world")); span.set_attribute(KeyValue::new("key5", 123)); } span.end(); }); trace_benchmark_group(c, "span-creation-span-builder", |tracer| { // This is similar to the simple span creation, but allows // attributes and other properties to be set during creation. // It is slightly slower than the simple span creation due to the fact that // attributes are collected into a vec! and allocated, even before sampling // decision is made. let mut span = tracer .span_builder("span-name") .with_attributes([ KeyValue::new("key1", false), KeyValue::new("key2", "hello"), KeyValue::new("key3", 123.456), ]) .start(tracer); if span.is_recording() { span.set_attribute(KeyValue::new("key4", "world")); span.set_attribute(KeyValue::new("key5", 123)); } span.end(); }); trace_benchmark_group(c, "span-creation-tracer-in-span", |tracer| { // This is similar to the simple span creation, but also does the job of activating // the span in the current context. // It is slower than other approaches of activation due to the fact that // context activation is done, irrespective of sampling decision. tracer.in_span("span-name", |ctx| { let span = ctx.span(); if span.is_recording() { span.set_attribute(KeyValue::new("key1", false)); span.set_attribute(KeyValue::new("key2", "hello")); span.set_attribute(KeyValue::new("key3", 123.456)); span.set_attribute(KeyValue::new("key4", "world")); span.set_attribute(KeyValue::new("key5", 123)); } }); }); trace_benchmark_group(c, "span-creation-tracer-in-span-with-builder", |tracer| { // This is similar to the simple span creation, but also does the job of activating // the span in the current context. tracer.in_span_with_builder( SpanBuilder::from_name("span-name").with_attributes([ KeyValue::new("key1", false), KeyValue::new("key2", "hello"), KeyValue::new("key3", 123.456), ]), |ctx| { let span = ctx.span(); if span.is_recording() { span.set_attribute(KeyValue::new("key4", "world")); span.set_attribute(KeyValue::new("key5", 123)); } }, ); }); trace_benchmark_group(c, "span-creation-simple-context-activation", |tracer| { // This optimizes by bypassing the context activation // based on sampling decision, and hence it is faster than the // tracer.in_span approach. let mut span = tracer.start("span-name"); if span.is_recording() { span.set_attribute(KeyValue::new("key1", false)); span.set_attribute(KeyValue::new("key2", "hello")); span.set_attribute(KeyValue::new("key3", 123.456)); let _guard = mark_span_as_active(span); Context::map_current(|cx| { let span_from_context = cx.span(); span_from_context.set_attribute(KeyValue::new("key4", "world")); span_from_context.set_attribute(KeyValue::new("key5", 123)); }); } }); trace_benchmark_group( c, "span-creation-span-builder-context-activation", |tracer| { // This optimizes by bypassing the context activation // based on sampling decision, and hence it is faster than the // tracer.in_span approach. let span = tracer .span_builder("span-name") .with_attributes([ KeyValue::new("key1", false), KeyValue::new("key2", "hello"), KeyValue::new("key3", 123.456), ]) .start(tracer); if span.is_recording() { let _guard = mark_span_as_active(span); Context::map_current(|cx| { let span_from_context = cx.span(); span_from_context.set_attribute(KeyValue::new("key4", "world")); span_from_context.set_attribute(KeyValue::new("key5", 123)); }); } }, ); let provider = sdktrace::SdkTracerProvider::builder() .with_sampler(sdktrace::Sampler::AlwaysOn) .with_simple_exporter(VoidExporter) .build(); let tracer = provider.tracer("active-parent"); let parent_cx = Context::new().with_span(tracer.start("parent")); c.bench_function("span-creation-with-active-parent", |b| { b.iter(|| { let mut span = tracer.start_with_context("span-name", &parent_cx); span.end(); }); }); } #[derive(Debug)] struct VoidExporter; impl SpanExporter for VoidExporter { async fn export(&self, _spans: Vec) -> OTelSdkResult { Ok(()) } } fn trace_benchmark_group(c: &mut Criterion, name: &str, f: F) { let mut group = c.benchmark_group(name); group.bench_function("always-sample", |b| { let provider = sdktrace::SdkTracerProvider::builder() .with_sampler(sdktrace::Sampler::AlwaysOn) .with_simple_exporter(VoidExporter) .build(); let always_sample = provider.tracer("always-sample"); b.iter(|| f(&always_sample)); }); group.bench_function("never-sample", |b| { let provider = sdktrace::SdkTracerProvider::builder() .with_sampler(sdktrace::Sampler::AlwaysOff) .with_simple_exporter(VoidExporter) .build(); let never_sample = provider.tracer("never-sample"); b.iter(|| f(&never_sample)); }); group.finish(); } #[cfg(all(not(target_os = "windows"), feature = "bench_profiling"))] criterion_group! { name = benches; config = Criterion::default() .warm_up_time(std::time::Duration::from_secs(1)) .measurement_time(std::time::Duration::from_secs(2)) .with_profiler(PProfProfiler::new(100, Output::Flamegraph(None))); targets = criterion_benchmark } #[cfg(any(target_os = "windows", not(feature = "bench_profiling")))] criterion_group! { name = benches; config = Criterion::default() .warm_up_time(std::time::Duration::from_secs(1)) .measurement_time(std::time::Duration::from_secs(2)); targets = criterion_benchmark } criterion_main!(benches);