Skip to main content
Beside the ConnectRPC services, the server’s listener (ADDR, default :8780) serves a few plain HTTP routes. They are registered in cmd/server/main.go and implemented in internal/server.

POST /functions/{function}/invoke

FunctionsService.Invoke over plain HTTP. The request body is the input and the response body is the output, both in the contract’s encoding, with no envelope. With Prefer: respond-async it is FunctionsService.InvokeAsync instead. Only POST is routed. Other methods on this path fall through to the web UI.

Path and query

Request headers

A browser cannot send these content types cross-site without a CORS preflight, which the server never grants, so a session cookie alone cannot invoke from another origin. The body may be at most 64 MiB. The call waits for its result for up to INVOKE_DEADLINE (default 5 minutes). Plain HTTP has no way to set a shorter or longer deadline. Use Invoke with Connect-Timeout-Ms for that. Disconnecting cancels the wait.

Responses

A FunctionError body uses the protobuf JSON names:

Response headers

On a sync call that ran (200, 500 or 504): On 202:

Location

The Location on a 202 names /invocations/<id>, which GET /invocations/{id} below serves.

Error body

Every failure that is not a function’s own error answers application/json:
code is the Connect code name. violations appears only when the input broke the contract, one ContractViolation per failure.

Status codes

The status follows the Connect protocol’s mapping for the code. The checks run in this order: authentication, the version query, the function and version lookup, Alchemy-Retries, Content-Type, then the body.

GET /invocations/{id}

FunctionsService.GetInvocation over plain HTTP. It answers 200 with the invocation as JSON, the same message GetInvocation returns, while it waits, runs and after it ends. An unknown id, or one in another tenant, answers 404 with the error body above.

GET /healthz

Answers 204 with no body while the process serves HTTP. It checks no dependency. It is not authenticated.

GET /scaling

The demand KEDA’s metrics-api scaler sizes the host Deployment on (host.autoscaling in the Helm chart). Every server gives the same answer, since the queue depth is read from NATS. It is not authenticated.
If the queue cannot be read, it answers 503 with the error as plain text. See Autoscaling.

The web UI

Every other path is the web UI. A binary built with -tags embedui serves ui/build with a fallback to index.html for client routes. Without the tag it serves a placeholder page that says how to build the UI.