span size perf test

Author: realark

perf-tests/build.gradle

@@ -0,0 +1,65 @@
+plugins {
+  id 'java'
+}
+
+java {
+  toolchain {
+    languageVersion = JavaLanguageVersion.of(17)
+    vendor = JvmVendorSpec.ADOPTIUM
+  }
+}
+
+repositories {
+  mavenCentral()
+}
+
+def langchainVersion = '1.8.0'
+
+ext {
+  otelVersion = rootProject.ext.otelVersion
+  junitVersion = rootProject.ext.junitVersion
+  slf4jVersion = rootProject.ext.slf4jVersion
+}
+
+dependencies {
+  testImplementation project(":braintrust-sdk")
+  testImplementation project(":braintrust-sdk:instrumentation:langchain_1_8_0")
+  testImplementation project(":braintrust-java-agent:instrumenter")
+  testImplementation(testFixtures(project(":test-harness")))
+
+  testImplementation "io.opentelemetry:opentelemetry-api:${otelVersion}"
+  testImplementation "io.opentelemetry:opentelemetry-sdk:${otelVersion}"
+  testImplementation "io.opentelemetry:opentelemetry-sdk-trace:${otelVersion}"
+  testImplementation "io.opentelemetry:opentelemetry-sdk-logs:${otelVersion}"
+  testImplementation "io.opentelemetry:opentelemetry-sdk-metrics:${otelVersion}"
+
+  testImplementation "dev.langchain4j:langchain4j:${langchainVersion}"
+  testImplementation "dev.langchain4j:langchain4j-http-client:${langchainVersion}"
+  testImplementation "dev.langchain4j:langchain4j-open-ai:${langchainVersion}"
+
+  testImplementation 'net.bytebuddy:byte-buddy-agent:1.17.5'
+
+  testImplementation "org.junit.jupiter:junit-jupiter:${junitVersion}"
+  testRuntimeOnly 'org.junit.platform:junit-platform-launcher'
+  testRuntimeOnly "org.slf4j:slf4j-simple:${slf4jVersion}"
+}
+
+// Disable the default test task so perf tests don't run during ./gradlew test or check.
+// Run explicitly with: ./gradlew :perf-tests:perfTest
+test {
+  enabled = false
+}
+
+task perfTest(type: Test) {
+  useJUnitPlatform()
+  workingDir = rootProject.projectDir
+
+  // Performance tests can be slow — give them generous timeouts
+  systemProperty 'junit.jupiter.execution.timeout.default', '120s'
+
+  testLogging {
+    events "passed", "skipped", "failed"
+    showStandardStreams = true
+    exceptionFormat "full"
+  }
+}

perf-tests/src/test/java/dev/braintrust/perf/PerfResult.java

@@ -0,0 +1,27 @@
+package dev.braintrust.perf;
+
+/**
+ * Captures the result of a single performance test run.
+ *
+ * @param config the configuration that produced this result
+ * @param payloadBytes total bytes of the OTLP HTTP request body captured at the wire
+ * @param spanCount number of spans that were exported (as observed by the server)
+ * @param requestCount number of HTTP requests received by the capture server
+ */
+public record PerfResult(PerfRunConfig config, long payloadBytes, int spanCount, int requestCount) {
+
+    /** Bytes per span (approximate). */
+    public double bytesPerSpan() {
+        return spanCount > 0 ? (double) payloadBytes / spanCount : 0;
+    }
+
+    public double payloadMB() {
+        return payloadBytes / (1024.0 * 1024.0);
+    }
+
+    public String summary() {
+        return String.format(
+                "[%s] %d request(s), %d span(s), %.3f MB total (%.1f bytes/span)",
+                config.name(), requestCount, spanCount, payloadMB(), bytesPerSpan());
+    }
+}

perf-tests/src/test/java/dev/braintrust/perf/PerfRunConfig.java

@@ -0,0 +1,24 @@
+package dev.braintrust.perf;
+
+/**
+ * Describes the configuration for a single performance test run.
+ *
+ * <p>Each field controls the shape of the multi-turn conversation that will be generated and
+ * exported. Add new fields here as you add new scenarios (e.g. tool use, streaming, etc.).
+ *
+ * @param name human-readable label for this configuration
+ * @param turns number of conversational turns (user messages sent to the AI service)
+ * @param includeImageAttachment whether to include an image attachment in the first user message
+ */
+public record PerfRunConfig(String name, int turns, boolean includeImageAttachment) {
+
+    /** A multi-turn conversation with an image attachment on the first turn. */
+    public static PerfRunConfig multiTurnWithAttachment() {
+        return new PerfRunConfig("multi-turn-with-attachment", 10, true);
+    }
+
+    /** A multi-turn conversation with text only (no attachments). */
+    public static PerfRunConfig multiTurnTextOnly() {
+        return new PerfRunConfig("multi-turn-text-only", 3, false);
+    }
+}