Lambda ADOT vs Powertools: Tracing Tradeoffs on Node 20 Runtimes

On Node 20, the wrong tracing stack adds tens of milliseconds to INIT and clutters X-Ray with noise while missing the one Dynamo span you needed. Choose ADOT when you need OpenTelemetry fan-out to multiple backends; choose Powertools Tracer when you want lean X-Ray annotations with minimal cold-start tax.

⚡ TL;DR: Powertools Tracer ≈ thin X-Ray SDK wrapper, fast INIT, great for Lambda-centric ops; ADOT layer ≈ OTel collector sidecar semantics, heavier INIT, better for multi-backend / existing OTel. Sample aggressively at high RPS; never trace every warm invoke at 100%. Pair with Lambda Powertools structured logs and Lambda Cold Starts on Node 20.

Cold-start and package cost

// powertools-tracer.ts — lean default for Node 20
import { Tracer } from "@aws-lambda-powertools/tracer";
import { captureLambdaHandler } from "@aws-lambda-powertools/tracer/middleware";
import middy from "@middy/core";

const tracer = new Tracer({ serviceName: "checkout" });

export const handler = middy(async (event) => {
  const segment = tracer.getSegment();
  const sub = segment?.addNewSubsegment("dynamo.PutItem");
  try {
    // AWS SDK v3 clients auto-capture when tracer.captureAWSv3Client(client)
    await putOrder(event);
  } finally {
    sub?.close();
  }
  return { ok: true };
}).use(captureLambdaHandler(tracer));

// adot — only when you need OTLP export (illustrative env)
// Layers: arn:aws:lambda:region:account:layer:aws-otel-nodejs-*-ver-*:1
// Env: AWS_LAMBDA_EXEC_WRAPPER=/opt/otel-handler
//     OTEL_PROPAGATORS=tracecontext,baggage,xray
//     OTEL_NODE_DISABLED_INSTRUMENTATIONS=fs,dns

❌ Enabling every OTel instrumentation (fs, dns, net) on a 128 MB function — INIT balloons and spans become unreadable.

Sampling at high RPS

At thousands of RPS, 100% tracing is a tax. Use reservoir + rate sampling; keep errors and slow traces.

{
  "version": 2,
  "rules": [
    {
      "description": "checkout errors and p99",
      "host": "*",
      "http_method": "*",
      "url_path": "/checkout*",
      "fixed_target": 2,
      "rate": 0.01
    }
  ],
  "default": { "fixed_target": 1, "rate": 0.005 }
}

Powertools: put annotations (orderId, tenant) not megabyte metadata. ADOT: drop high-cardinality attributes in a collector processor before export.

Decision guide

Need only X-Ray + Powertools Logger/Metrics already? → Powertools Tracer
Need Jaeger/Tempo/Honeycomb + Lambda?             → ADOT / OTel
Sub-100ms p99 tool Lambda?                          → Prefer Powertools or no tracer on warm path
Multi-language estate with OTel conventions?        → ADOT for consistency

Closing checklist

✅ Dos
– ✅ Measure INIT with and without the tracing layer before standardizing
– ✅ Sample; annotate IDs; capture AWS SDK clients explicitly
– ✅ Disable noisy OTel instrumentations
– ✅ Correlate with Powertools Logger correlationId
– ✅ Document the choice in the service README

❌ Don’ts
– ❌ Don’t run ADOT + full X-Ray SDK double instrumentation
– ❌ Don’t 100%-sample hot synchronous APIs
– ❌ Don’t put PII in annotations/metadata
– ❌ Don’t ignore layer version pins in IaC

Related reading

• Lambda Powertools Node Structured Logs
• Lambda Cold Starts Node 20
• AWS Lambda Extensions API Custom Monitoring
• Lambda Warm Pools

Last updated on September 11, 2026