Observability

The runtime exposes diagnostics, a bounded observer stream, and an optional OpenTelemetry package. These surfaces contain no credentials or payloads.

Diagnostics#

Copydiagnostics = client.diagnostics()
print(diagnostics.in_flight_operations)
print(diagnostics.dropped_observer_events)
print(diagnostics.observer_failures)

RuntimeDiagnostics also counts concurrency and byte-budget rejections, response limit and close failures, scheduler and executor overload, telemetry failures, and lifecycle state. Each call returns an immutable snapshot.

Observers#

Copyfrom repost_client import ClientOptions, ObserverEvent, ObserverEventKind
from repost_sdk import RepostClient
 
def observe(event: ObserverEvent) -> None:
    if event.kind is ObserverEventKind.OPERATION_END:
        record_end(event.operation_id, event.delivery_state)
 
client = RepostClient(ClientOptions(observer=observe))

Observers receive operation start/end, attempt start/end, retry-delay, and cancellation events. They run on a dedicated dispatch thread, never the caller or transport thread. The queue holds 1,024 events. When it fills, the runtime drops and counts events instead of slowing a send. A slow or throwing observer cannot change the send result.

OpenTelemetry#

Install the isolated bridge package and wire its two explicit adapters:

Copypython -m pip install repost-client-otel

Copyfrom repost_client import ClientOptions
from repost_client_otel import metrics_observer, telemetry
from repost_sdk import RepostClient
 
client = RepostClient(ClientOptions(
    observer=metrics_observer(),
    telemetry=telemetry(),
))

The metrics observer records exactly these instruments:

Metrics use the low-cardinality attributes outcome, error.code, delivery.state, and http.status.class. Tracing emits one repost.send span and one repost.send.attempt child per attempt. The attempt carries http.request.method, repost.retry.attempt, network.protocol.name, and a success status when known. Trace context is propagated with traceparent.

Continue#