How telemetry is sent
You're contributing to the SDK, or wiring up a collector, and want to know why a collector that's down never slows a page. This page shows the wire format, how metrics are shaped, and how batches leave a server. What's sent and how to configure it is in Telemetry.
The wire format
Telemetry is OpenTelemetry over OTLP/HTTP, encoded as JSON, posted to the standard paths under the configured endpoint: /v1/traces, /v1/metrics and /v1/logs. JSON keeps a protobuf library out of the edge bundle. Bodies are gzipped. A failed send is retried only on 429, 502, 503 and 504; any other error drops the batch.
Metrics
Metric names follow the OpenTelemetry semantic conventions, such as http.client.request.duration for upstream calls. The SDK records no http.server.* metrics: the requests your site serves are your framework's to measure. Metrics use delta temporality: each batch carries what happened since the last one, because a Cloudflare Workers isolate can disappear at any moment and has no long-lived counter to report. Prometheus-style backends need the collector's deltatocumulative processor to turn them into cumulative counters.
Every batch carries these resource attributes:
| Attribute | Value |
|---|---|
service.name | OTEL_SERVICE_NAME, else the site ID, else decocms-site |
service.version | The deployed commit, read from the first variable your host sets (DECO_COMMIT_SHA, WORKERS_CI_COMMIT_SHA, CF_PAGES_COMMIT_SHA, VERCEL_GIT_COMMIT_SHA, GITHUB_SHA, RENDER_GIT_COMMIT, SOURCE_VERSION, COMMIT_SHA), else unknown |
deployment.environment.name | VERCEL_ENV when set, else development when NODE_ENV is development, else production |
deco.site | The site ID, when there is one |
deco.release | The revision being served |
The standard OTEL_RESOURCE_ATTRIBUTES variable overrides any of them, for example deployment.environment.name=preview.
Sending in the background
Measurements are aggregated in memory and sent in batches, never in front of a response. On Cloudflare Workers, which run no timers between requests, a batch goes out after the response, inside ctx.waitUntil. Elsewhere, it goes out on a timer that doesn't keep the process alive. A collector that's down or slow loses that one batch; nothing waits for it and nothing piles up.
Scrubbing
The SDK removes cookie, authorization and similar headers, tokens and request and response bodies before anything is encoded. Scrubbing happens on the sending side, not in a collector, so a third-party endpoint gets the same guarantee as the hosted Deco CMS.
Several CMS instances
The instrumented fetch reports its measurements to the CMS in the same process, which sends them where its telemetry option points. The instrumented fetch reports to the most recently created CMS with telemetry; several sites in one process share one destination.
Analytics events
Analytics is separate from telemetry: it doesn't use the telemetry option or the OTLP path above. The script AnalyticsScript renders sends page views in the One Dollar Stats tracker's wire format: a JSON object with the page URL (without query string) and a list of events, each with its type, the referrer when it's from another site, and its properties. When the encoded payload is short, it goes as a GET with the JSON, base64-encoded, in the ?data= parameter; otherwise it's sent with navigator.sendBeacon or a POST.