@davidilie/telemetry-node

OpenTelemetry-backed events, logs, measurements, exceptions, and spans for Node.js. Use the /register entrypoint in a standalone process. Use the root entrypoint when a framework already owns the OpenTelemetry provider.

Node.js 20 or newer is supported. The package publishes ESM and CommonJS.

pnpm add @davidilie/telemetry-node @opentelemetry/api

Standalone Node.js

Import registration before modules that need to be auto-instrumented.

import { registerNodeTelemetry } from "@davidilie/telemetry-node/register";

const deployedSha = process.env.GIT_SHA!;
const telemetry = registerNodeTelemetry({
  resource: {
    serviceName: "email-worker",
    serviceVersion: deployedSha,
    environment: process.env.DEPLOYMENT_ENV ?? "production",
    namespace: "notifications",
    repositoryUrl: "https://github.com/example/email-worker",
    commitSha: deployedSha,
    attributes: {
      "davidapps.project.id": "notifications",
    },
  },
});

await telemetry.withSpan(
  "jobs.process",
  async () => {
    telemetry.capture("job.started", { "messaging.destination.name": "email" });
  },
  { "job.type": "email" },
);

process.once("SIGTERM", async () => {
  await telemetry.shutdown();
  process.exit(0);
});

registerNodeTelemetry installs one NodeSDK per JavaScript global. Repeated calls return the first client and ignore later configuration until that client successfully shuts down. Do not register this package beside another provider.

The optional sdk property accepts NodeSDKConfiguration fields other than resource and serviceName. Use it for maintained instrumentations, readers, processors, or exporters when environment-based defaults are insufficient.

Export configuration

The standalone SDK follows standard OpenTelemetry environment variables. For Alloy's private OTLP/HTTP receiver:

OTEL_SERVICE_NAME=email-worker
OTEL_EXPORTER_OTLP_ENDPOINT=http://alloy.observability.svc.cluster.local:4318
OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf

Set OTEL_TRACES_EXPORTER=none, OTEL_LOGS_EXPORTER=none, or OTEL_METRICS_EXPORTER=none to disable a signal. Configure resource data in the JavaScript resource object above; it remains the package's canonical mapping.

Server workloads should use private Alloy directly. Do not put the public browser/mobile routing key on a server unless the workload genuinely runs outside the cluster and the gateway has intentionally enabled that signal.

Existing provider

Frameworks such as Next.js may already own provider registration. Create only the client from the runtime-light root entrypoint:

import { createNodeTelemetry } from "@davidilie/telemetry-node";

const telemetry = createNodeTelemetry({
  resource: {
    serviceName: "web",
    serviceVersion: process.env.GIT_SHA,
    environment: process.env.DEPLOYMENT_ENV,
    commitSha: process.env.GIT_SHA,
    attributes: { "davidapps.project.id": "web" },
  },
});

Without a registered provider, trace and metric API calls are no-ops. The structured-console log path still works.

Client configuration

| Option | Behavior | | --- | --- | | resource | Required shared service/project/exact-release identity | | enabled / consent | Gate future client-created signals; pending/denied do not queue | | sampleRate | 0..1; standalone registration applies it once to root traces and logs, while metrics remain aggregated rather than randomly sampled | | beforeSend | Async transform/drop hook with a second sanitization pass | | debug / onError | Contained telemetry diagnostics | | instrumentationName / instrumentationVersion | OTel tracer/logger/meter library identity; not application release identity | | structuredConsole | JSON stdout/stderr fallback, default true | | measurementMode | spans, metrics, or both | | metricAttributeAllowlist | Bounded signal keys allowed on metric points | | maxMetricInstruments | Maximum histogram name/unit pairs, default 64 | | sdk | /register only: provider configuration excluding resource/service name |

Do not put the deployed commit into instrumentationVersion; it belongs in resource serviceVersion/commitSha. The instrumentation version identifies the instrumentation library itself.

Signal mapping

| Core call | Node behavior | | --- | --- | | capture(name, attrs) | Adds an event to the active span; otherwise creates and ends a short internal span containing the event | | captureException(error, attrs) | Records on the active span and marks it error; otherwise creates an exception span | | log(level, message, attrs) | Emits through the OTel Logs API and, by default, one JSON line to stdout/stderr | | measure(name, value, attrs, unit) | By default records a short measurement:<name> trace span; can instead/additionally record an OTel histogram | | withSpan(name, operation, attrs, options) | Creates an active span with an optional portable kind, records/rethrows application failures, and always ends it |

Measurements default to spans so they remain observable without an OTel metric reader. Configure the representation explicitly when needed:

const telemetry = registerNodeTelemetry({
  resource,
  measurementMode: "both", // "spans" (default), "metrics", or "both"
  metricAttributeAllowlist: ["checkout.provider"],
  maxMetricInstruments: 32,
  sdk: {
    metricReaders: [reader],
  },
});

When /register sees configured sdk.metricReaders, its default mode becomes metrics; otherwise it is spans. A framework client chooses the mode supplied by the caller. Histogram names must match the adapter's bounded instrument-name syntax, and no more than 64 name/unit pairs are created by default. Only explicitly allowlisted signal attributes reach metric points; resource labels come from the provider. Metrics still require a reader/exporter.

Structured logs and correlation

JSON console output is enabled by default so Kubernetes log collection reaches VictoriaLogs even when the framework provider did not configure OTel Logs. It includes service_name, service_version, and active trace_id/span_id.

const telemetry = createNodeTelemetry({
  resource,
  structuredConsole: false, // OTel Logs is the sole desired route.
});

The default can duplicate a logical log if both stdout and OTLP Logs are collected into the same backend. Disable one route deliberately.

Spans and context

The NodeTelemetryClient overrides the generic core withSpan so the new span is active during the operation. Use it when events/logs need the same trace. client.startSpan() creates a span but does not make it active.

Both methods accept { kind: "internal" | "server" | "client" | "producer" | "consumer" } after attributes. Manual span handles also expose setStatus("unset" | "ok" | "error", message?).

The package also exports provider-level helpers:

import {
  currentTraceContext,
  recordException,
  startSpan,
  withSpan,
} from "@davidilie/telemetry-node";

These use the globally registered provider and do not add a client's resource mapping, context, or beforeSend. They sanitize their direct names, attributes, exceptions, and status text, but do not consult a root client's per-call consent/sampling; only the globally registered provider's policy can gate them. Use the initialized client for application signals and reserve these helpers for reviewed provider integrations. withSpan establishes active context; startSpan does not.

Failure, consent, and lifecycle behavior

• Core capture calls isolate hook/export errors from the application.
• withSpan rethrows the application's failure after recording it.
• enabled: false, pending consent, and denied consent skip new client calls; they do not buffer for later.
• Standalone registration wraps a parent-based trace-ratio sampler in a mutable collection gate. setEnabled and setConsent update that provider-wide gate immediately, including for automatic instrumentation and sampled-parent spans. A custom sdk.sampler replaces the ratio sampler but remains inside the gate.
• The standalone client does not sample traces twice: the provider owns span, event, and exception trace decisions, while the adapter applies sampleRate once to logs. OTel histogram measurements are aggregated without random point sampling. A root-entrypoint client has no provider ownership, so its core per-call gate and an externally configured provider sampler can both apply; set one layer to 1 if multiplied rates are not intentional.
• State changes affect new work and do not erase an already exported or queued provider batch.
• flush() drains core work, but the Node adapter does not own a provider force-flush method. shutdown() on the /register client shuts down the NodeSDK and its processors.
• Only a client returned by /register owns and shuts down the installed NodeSDK. A client created from the root entrypoint does not own the framework provider.
• onError(error, context) receives contained hook/adapter/span/lifecycle failures. Its own failure is ignored; never recurse into the same client.

See the core API, privacy rules, and troubleshooting guide.