Telemetry events emitted by Candil during LLM inference.
All events are prefixed with [:candil, ...] so a single handler
can attach to everything Candil does.
Events
[:candil, :inference, :start]— a chat/embedding/streaming request begins. Metadata::model,:provider,:kind(:chat | :embed | :stream),:request_id.[:candil, :inference, :stop]— a request completes successfully. Metadata adds:duration_ms,:tokens_in,:tokens_out.[:candil, :inference, :token]— a streamed token arrives (only fires forchat_stream). Metadata::request_id,:tokens_so_far.[:candil, :inference, :error]— a request fails. Metadata::model,:provider,:kind,:request_id,:reason,:duration_ms.[:candil, :cost, :estimate]— a cost estimate was computed. Metadata::provider,:model,:tokens_in,:tokens_out,:cost_usd.[:candil, :cancellation]— a generation was cancelled. Metadata::request_id,:reason.
Measurements
Measurements follow the duration_in_native_time convention from
:telemetry. To get milliseconds, use
System.convert_time_unit(duration, :native, :millisecond).
Example
:telemetry.attach("candil-logger", [:candil, :inference, :stop], fn _name, measurements, meta, _ ->
Logger.info("inference done in #{measurements.duration_ms}ms")
end, nil)Summary
Functions
Emit a :cancellation event.
Emit a :cost event.
Emit an :error event.
Emit a :start event.
Emit a :stop event with timing data.
Emit a :token event for streaming responses.
Types
Functions
Emit a :cancellation event.
@spec emit_cost(atom(), String.t(), non_neg_integer(), non_neg_integer(), float()) :: :ok
Emit a :cost event.
@spec emit_error(String.t(), event_kind(), non_neg_integer(), atom(), map()) :: :ok
Emit an :error event.
@spec emit_start(String.t(), event_kind(), map()) :: :ok
Emit a :start event.
@spec emit_stop(String.t(), event_kind(), non_neg_integer(), keyword() | map()) :: :ok
Emit a :stop event with timing data.
@spec emit_token(String.t(), non_neg_integer()) :: :ok
Emit a :token event for streaming responses.