OpenTelemetry

Native distributed tracing through @opentelemetry/api, with zero setup when no SDK is registered.

@opentelemetry/api is a required peer dependency of @zap-studio/fetch. It’s a tiny, side-effect-free package that’s a no-op until an app registers a real SDK, so installing it costs nothing at runtime for consumers who never set one up — the same deal logger? already gives.

npm install @opentelemetry/api

yarn add @opentelemetry/api

pnpm add @opentelemetry/api

bun add @opentelemetry/api

deno add npm:@opentelemetry/api

There’s no option to configure — tracing is always on, and does nothing observable until your app registers an OpenTelemetry SDK (a Datadog agent, Honeycomb, Vercel OTel, a self-hosted collector, etc.):

import { createFetch } from "@zap-studio/fetch";

const { api } = createFetch({ baseURL: "https://api.example.com" });

await api.get("/users/1", UserSchema);

What Gets Instrumented

Every request produces a CLIENT span named after the HTTP method (e.g. GET), with:

The span’s trace context is injected into the outgoing request’s headers (traceparent, and tracestate if your propagator uses it), so the call continues the caller’s distributed trace instead of starting a new one.

On failure — a non-2xx response, or a thrown error (network failure, aborted request, JSON parse failure, validation failure) — the span’s status is set to ERROR. A thrown error is additionally recorded as a span exception via recordException.

Edge Runtime Caveat

Implicit context propagation — auto-linking the “active span” across await boundaries so your own spans nest correctly under the request span — depends on the host runtime supporting AsyncLocalStorage/async_hooks. Some edge runtimes don’t support this reliably.

Span creation and attribute-setting always work everywhere, regardless of this. Only the automatic parent/child linking across async boundaries can be affected on those runtimes.

Log Correlation

Pair this with logger: once a span is active, ConsoleLogger automatically stamps trace_id/span_id onto every log line, so a request’s spans and logs line up in whatever backend you send them to.