Review Kafka and RabbitMQ attributes

[lmolkova]

Based on the messaging SIG discussion on 2/15/24, we should target stabilizing Kafka and RabbitMQ semconv along with the messaging semconv stability.

While we can always add span attributes, but adding attributes to the metrics is breaking #722.

So prior to stability, we should

• review if some important (for metrics) attributes are missing
• have kafka and rabbitmq metrics instrumentation prototypes

Update: see https://github.com/open-telemetry/semantic-conventions/pull/798#discussion_r1516427699

• we should also check rabbitmq/kafka attributes to see if any of the existing ones are good candidates to be generalized

[lmolkova]

Dropping some initial research results for Kafka here.

TL;DR: looking into kafka JMX metrics, 'native' metrics proposal, javadocs for client APIs, I can see the following properties that might be interesting on metrics, have low cardinality, no sensitive information, apply on the operation level (rather than transport).

• node.id (network.peer.address? The IP address the final response was received from, if known on the logical level)
• client.rack (like an availability zone and maybe fits better as a resource/scope attribute)
• enable.auto.commit - receive mode

Others like client.id (messaging.client_id) would be controversial.

---

Producer

Documented in 'native' kafka metrics:

• client.id
• transactional.id - per-client-instance configuration that plays role in transaction support.

These are per-client instance and potentially have high cardinality. Reporting them on OTel metrics is controversial.

Consumer

Documented in 'native' kafka metrics:

• client.rack. Availability zone-equivalent
• group.id - same as messaging.kafka.consumer.group
• group_instance_id, group_member_id - describe client/machine instances and don't fit otel metrics due to cardinality (partially covered by service.instance.id)

Consumer config

• enable.auto.commit. There are similar modes in other messaging systems when messages are settled upon successful delivery. Perhaps we can think if it's useful on metrics

Links:

1. Kafka 'native' metrics https://cwiki.apache.org/confluence/display/KAFKA/KIP-714%3A+Client+metrics+and+observability
2. Java API docs
   • KafkaProducer and ProducerRecord,
   • KafkaConsumer and ConsumerRecord
3. Kafka configuration options for producers and consumers
4. Kafka JMX metrics

[lmolkova]

adding research on RabbitMQ

TL;DR: pre-stability:

• [ ] we should consider adding redelivered and delivery_mode attributes for RabbitMQ.
• [x] we should consider overriding span names for rabbitmq to follow rabbitmq terminology when it does not align (consume, get, etc)

can be added incrementally:

• we should consider adding messaging.header.* template similar to http headers.
• we should consider adding general-purpose attribute to record time-in-queue on spans (and generic metric). It's extremely useful and very common.

---

Java otel instr reports

• rabbitmq.record.queue_time_ms - time in queue for message
• rabbitmq.command (basic.publish, basic.get, queue.create, exchange.declare etc) - perhaps should be changed to messaging.operation rabbitmq.queue - perhaps should be changed to messaging.destination.name
• rabbitmq.delivery_mode - https://www.rabbitmq.com/consumers.html#message-properties
• messaging.header.* template

Python otel doesn't report anything new

RabbitMQ message properties that could be interesting on metrics (low cardinality, logical level):

• Redelivered bool flag is message was delivered before
• Delivery mode (persistent vs transient message)
• type - Application-specific message type, e.g. "orders.created"
• Content type and Content encoding - e.g. "application/json" and "gzip". Used by applications, not core RabbitMQ

Collector rabbitmq scraper does not report any attributes.

Client JMX metrics don't seem to have any attributes

[pyohannes]

we should also check rabbitmq/kafka attributes to see if any of the existing ones are good candidates to be generalized

Based on discussions in #798, we should also look into standardizing messaging.kafka.message.offset.

[pyohannes]

Discussed in the messaging workgroup: we'll need a RabbitMQ-specific extension for metrics because the routing key is needed on RabbitMQ metrics.

[pyohannes]

This issue is related: https://github.com/open-telemetry/semantic-conventions/issues/1156