Telemetry
AuthProxy emits OpenTelemetry traces and metrics under the instrumentation scope Cratis.AuthProxy,
and exports redacted logs through the shared setup.
It is off until you point it somewhere: with no OTEL_* variable set nothing is exported and the proxy
behaves exactly as before.
Setup uses the shared Cratis.OpenTelemetry package, so the standard OpenTelemetry environment variables
apply and need no AuthProxy-specific configuration.
Turning it on
Section titled “Turning it on”| Variable | Effect |
|---|---|
OTEL_EXPORTER_OTLP_ENDPOINT |
Where to send traces, metrics and logs over OTLP. Unset means nothing is exported. |
OTEL_EXPORTER_OTLP_HEADERS |
Headers for the exporter, such as an API key. |
OTEL_EXPORTER_OTLP_PROTOCOL |
grpc or http/protobuf. |
OTEL_SERVICE_NAME |
The service.name of the proxy. Defaults to cratis-authproxy. |
OTEL_RESOURCE_ATTRIBUTES |
Extra resource attributes, such as deployment.environment=prod. |
Azure Monitor and Application Insights accept OTLP through an OpenTelemetry Collector (or the Azure Monitor
collector exporter): point OTEL_EXPORTER_OTLP_ENDPOINT at the collector and let it forward to Application
Insights.
Trace context
Section titled “Trace context”The W3C traceparent and tracestate headers a caller sends continue through the proxy to the service, so a
request is one trace from the caller to your application. The W3C baggage header and the legacy
Correlation-Context header are not forwarded: what a caller attaches to a request never reaches a service.
| Span | Covers |
|---|---|
cratis.authproxy.sign_in.notify |
Telling the application a sign-in completed. |
cratis.authproxy.identity.resolve |
Resolving the identity details of a signed-in caller (/.cratis/me). |
cratis.authproxy.tenant.resolve |
Resolving the tenant of a request. |
cratis.authproxy.invite.exchange |
Exchanging an invitation for a completed sign-in. |
cratis.authproxy.bearer.validate |
Validating the access token presented on a bearer route. |
Metrics
Section titled “Metrics”| Metric | Unit | Attribute | Counts |
|---|---|---|---|
cratis.authproxy.sign_ins |
{sign_in} |
cratis.authproxy.sign_in.outcome |
Completed sign-ins. |
cratis.authproxy.identity.denials |
{denial} |
cratis.authproxy.identity.denial_reason |
Identity verifications that refused the caller. |
cratis.authproxy.identity.cache.requests |
{request} |
cratis.authproxy.identity.cache.result |
Identity cache lookups, hit or miss. |
cratis.authproxy.bearer.validations |
{validation} |
cratis.authproxy.bearer.outcome |
Bearer token validations. |
cratis.authproxy.invite.exchanges |
{exchange} |
cratis.authproxy.invite.outcome |
Invitation exchanges. |
cratis.authproxy.upstream.errors |
{error} |
error.type |
Requests that could not be forwarded to a service. |
The names are constants on Cratis.AuthProxy.Telemetry.WellKnownTelemetryNames.
What is never recorded
Section titled “What is never recorded”Every attribute AuthProxy adds is a small, closed set of lowercase words. No token, claim value, subject, email, cookie, secret, URL or tenant identifier is ever added as a span name, span attribute, or metric attribute, so a telemetry backend never holds a credential or a person.
The standard ASP.NET Core and HTTP client instrumentation that comes with the shared setup also records requests, so AuthProxy redacts what it records before any exporter sees it:
| Where | Redacted |
|---|---|
| Incoming request spans | The query string is never recorded, so a sign-in callback’s code and state never appear. A path under /invite is recorded as /invite/{token}. The host (server.address, which is the tenant under subdomain tenancy) and the full URL are never recorded. |
| Outgoing request spans | Full URLs (url.full, http.url), paths, queries and destination hosts (server.address, http.host, net.peer.name, network.peer.address, dns.question.name) and destination ports are removed, including on failed calls. |
| Exceptions | exception.* span and event attributes and status descriptions are removed on every activity. Exception events, including native .NET HTTP client events, retain only their name and timestamp, not their payload. Span error status is kept. |
| Exported logs | Only metadata is exported: timestamp, severity, category, event ID and trace correlation. Every body and formatted message becomes Log details redacted; structured attributes, exceptions and scopes are excluded, at all log levels. This includes tenant-verification failures, bearer tenant refusals and Debug identity-endpoint URLs. The hosting, routing, Kestrel, reverse proxy and HTTP client loggers are still excluded below Warning. Local console messages retain their original details. |
| HTTP client metrics | Only http.request.method, http.response.status_code, error.type, network.protocol.version, url.scheme and http.connection.state are retained. The host the proxy called (server.address, dns.question.name) is not a metric attribute, because it can carry a tenant. |
Sanitization runs on completed activities before exporter queuing, not just during HTTP enrichment.
If path redaction throws, all span attributes are discarded; the span and its error status can still be
exported. Incoming url.full and http.url are explicitly removed. Standard HTTP method, response status,
timing and trace correlation remain available when redaction succeeds. Log event IDs and categories help
locate the original console diagnostics; exported logs deliberately do not include their detailed messages.
Resource attributes supplied through OTEL_RESOURCE_ATTRIBUTES are operator-controlled and are not sanitized;
do not put credentials or tenant identifiers in them.
The path of any other incoming request is recorded as it arrived. Paths that your application’s own routes use to carry a secret are your application’s to keep out of its URLs; to hide them, drop the request spans in your collector.