message-store-open-telemetry

OpenTelemetry distributed tracing for the message-store package (NATS JetStream event sourcing).

It wraps a message store and layers tracing on top of its public interface — the same interception pattern used in the JavaScript version. No tracing code lives in the message store core; all OpenTelemetry concerns are contained here.

• Publish → a PRODUCER span is started and W3C trace context is injected into the NATS message headers (not the event body, so application schemas stay clean).
• Consume → each handler runs inside a CONSUMER span whose parent is extracted from those headers, joining the same trace across the service hop. A backdated queue_wait span captures how long the message sat in the stream before delivery.
• ensure_stream, fetch and wait_for pass straight through, untouched.

Installation

pip install message-store-open-telemetry
# or: uv add message-store-open-telemetry

Depends on message-store>=0.4.0 (the version that exposes _raw_jetstream_message and the publish headers parameter this package relies on) and opentelemetry-api. nats-py arrives transitively through message-store.

Usage

Wrap the store immediately after building it:

import nats
from message_store import MessageStore
from message_store_open_telemetry import get_message_store_with_open_telemetry

nats_connection = await nats.connect("nats://127.0.0.1:4222")

message_store = get_message_store_with_open_telemetry(
    MessageStore(
        nats_connection,
        prefix="my-service",
        should_create_missing_streams=True,
    )
)

# Use `message_store` exactly as before — publishing and subscriptions are now traced.

The wrapper exposes the same surface as MessageStore. Type code that accepts either the concrete store or a wrapped one against MessageStoreProtocol (exported from message_store):

from message_store import MessageStoreProtocol

def register(store: MessageStoreProtocol) -> None:
    ...

Enabling the SDK

This library only uses the OpenTelemetry API. Spans are emitted only once the host process registers an OpenTelemetry SDK and tracer provider. With no SDK registered, every API call is a cheap no-op, so the wrapper is safe to apply in services that are not exporting traces yet.

A minimal SDK setup (configure the exporter for your collector):

from opentelemetry import trace
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter

provider = TracerProvider(resource=Resource.create({"service.name": "my-service"}))
provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
trace.set_tracer_provider(provider)

Emitted spans and attributes

Publish — nats publish <subject> (PRODUCER)

Attribute	Source
messaging.system	"nats"
messaging.operation	"publish"
messaging.destination.name	published subject
messaging.message.body.type	message.type
messaging.nats.stream	from the returned PubAck
messaging.message.seq	from the returned PubAck

Consume — nats process <subject> (CONSUMER)

Parented to (and linked to) the producer span when trace headers are present.

Attribute	Source
messaging.system	"nats"
messaging.operation	"process"
messaging.destination.name	message subject
messaging.message.seq	message sequence
messaging.message.body.type	message.type
messaging.nats.consumer	consumer (durable) name
messaging.nats.stream	raw JetStream message
messaging.nats.redelivery_count	raw JetStream message

queue_wait — a child span backdated to the moment JetStream stored the message and ended at delivery, so its duration is the stream dwell time (tagged with messaging.nats.dwell_ms and messaging.nats.redelivery_count).

Handler errors are recorded on the consumer span and re-raised, so the message store core keeps full ownership of ack/nak/term.

How context propagates

The publish seam injects the active context into a NATS headers dict via inject; the consume seam reads them back via extract. Whatever propagator your SDK installs (W3C tracecontext by default) is honoured — no propagator is forced here. Trace context never touches the persisted event payload.

Exports

• get_message_store_with_open_telemetry(store: MessageStoreProtocol) -> MessageStoreProtocol
• tracer — the library's Tracer instance, exported for advanced use.