Messaging systems

Status: Experimental

Definitions

Message

Although messaging systems are not as standardized as, e.g., HTTP, it is assumed that the following definitions are applicable to most of them that have similar concepts at all (names borrowed mostly from JMS):

A message is an envelope with a potentially empty payload. This envelope may offer the possibility to convey additional metadata, often in key/value form.

A message is sent by a message producer to:

• Physically: some message broker (which can be e.g., a single server, or a cluster, or a local process reached via IPC). The broker handles the actual delivery, re-delivery, persistence, etc. In some messaging systems the broker may be identical or co-located with (some) message consumers. With Apache Kafka, the physical broker a message is written to depends on the number of partitions, and which broker is the leader of the partition the record is written to.
• Logically: some particular message destination.

Messages can be delivered to 0, 1, or multiple consumers depending on the dispatching semantic of the protocol.

Producer

The “producer” is a specific instance, process or device that creates and publishes a message. “Publishing” is the process of sending a message or batch of messages to the intermediary or consumer.

Consumer

A “consumer” receives the message and acts upon it. It uses the context and data to execute some logic, which might lead to the occurrence of new events.

The consumer receives, processes, and settles a message. “Receiving” is the process of obtaining a message from the intermediary, “processing” is the process of acting on the information a message contains, “settling” is the process of notifying an intermediary that a message was processed successfully.

Intermediary

An “intermediary” receives a message to forward it to the next receiver, which might be another intermediary or a consumer.

Destinations

A destination is usually identified by some name unique within the messaging system instance, which might look like a URL or a simple one-word identifier. Traditional messaging, such as JMS, involves two kinds of destinations: topics and queues. A message that is sent (the send-operation is often called “publish” in this context) to a topic is broadcasted to all consumers that have subscribed to the topic. A message submitted to a queue is processed by a message consumer (usually exactly once although some message systems support a more performant at-least-once mode for messages with idempotent processing).

In a messaging system such as Apache Kafka, all destinations are topics. Each record, or message, is sent to a single consumer per consumer group. Consumer groups provide deliver once semantics for consumers of a topic within a group. Whether a specific message is processed as if it was sent to a topic or queue entirely depends on the consumer groups and their composition. For instance, there can be multiple consumer groups processing records from the same topic.

Message consumption

The consumption of a message can happen in multiple steps. First, the lower-level receiving of a message at a consumer, and then the logical processing of the message. Often, the waiting for a message is not particularly interesting and hidden away in a framework that only invokes some handler function to process a message once one is received (in the same way that the listening on a TCP port for an incoming HTTP message is not particularly interesting).

Conversations

In some messaging systems, a message can receive one or more reply messages that answers a particular other message that was sent earlier. All messages that are grouped together by such a reply-relationship are called a conversation. The grouping usually happens through some sort of “In-Reply-To:” meta information or an explicit conversation ID (sometimes called correlation ID). Sometimes a conversation can span multiple message destinations (e.g. initiated via a topic, continued on a temporary one-to-one queue).

Temporary destinations

Some messaging systems support the concept of temporary destination (often only temporary queues) that are established just for a particular set of communication partners (often one to one) or conversation. Often such destinations are unnamed or have an auto-generated name.

Conventions

Given these definitions, the remainder of this section describes the semantic conventions for Spans describing interactions with messaging systems.

Context propagation

A message may traverse many different components and layers in one or more intermediaries when it is propagated from the producer to the consumer(s). To be able to correlate consumer traces with producer traces using the existing context propagation mechanisms, all components must propagate context down the chain.

Messaging systems themselves may trace messages as the messages travels from producers to consumers. Such tracing would cover the transport layer but would not help in correlating producers with consumers. To be able to directly correlate producers with consumers, another context that is propagated with the message is required.

A message creation context allows correlating producers with consumers of a message and model the dependencies between them, regardless of the underlying messaging transport mechanism and its instrumentation.

The message creation context is created by the producer and should be propagated to the consumer(s). Consumer traces cannot be directly correlated with producer traces if the message creation context is not attached and propagated with the message.

A producer SHOULD attach a message creation context to each message. If possible, the message creation context SHOULD be attached in such a way that it cannot be changed by intermediaries.

This document does not specify the exact mechanisms on how the creation context is attached/extracted to/from messages. Future versions of these conventions will give clear recommendations, following industry standards including, but not limited to Trace Context: AMQP protocol and Trace Context: MQTT protocol once those standards reach a stable state.

Span name

The span name SHOULD be set to the message destination name and the operation being performed in the following format:

<destination name> <operation name>

The destination name SHOULD only be used for the span name if it is known to be of low cardinality (cf. general span name guidelines). This can be assumed if it is statically derived from application code or configuration. Wherever possible, the real destination names after resolving logical or aliased names SHOULD be used. If the destination name is dynamic, such as a conversation ID or a value obtained from a Reply-To header, it SHOULD NOT be used for the span name. In these cases, an artificial destination name that best expresses the destination, or a generic, static fallback like "(temporary)" for temporary destinations SHOULD be used instead.

The values allowed for <operation name> are defined in the section Operation names below. If the format above is used, the operation name MUST match the messaging.operation attribute defined for message consumer spans below.

Examples:

• shop.orders send
• shop.orders receive
• shop.orders process
• print_jobs send
• topic with spaces process
• AuthenticationRequest-Conversations process
• (temporary) send ((temporary) being a stable identifier for randomly generated, temporary destination names)

Span kind

A producer of a message should set the span kind to PRODUCER unless it synchronously waits for a response: then it should use CLIENT. The processor of the message should set the kind to CONSUMER, unless it always sends back a reply that is directed to the producer of the message (as opposed to e.g., a queue on which the producer happens to listen): then it should use SERVER.

Operation names

The following operations related to messages are defined for these semantic conventions:

Messaging attributes

[1]: If the message destination is either a queue or topic.

[2]: If value is true. When missing, the value is assumed to be false.

[3]: This should be the IP/hostname of the broker (or other network-level peer) this specific message is sent to/received from.

[4]: If different than inet and if any of net.sock.peer.addr or net.sock.host.addr are set. Consumers of telemetry SHOULD accept both IPv4 and IPv6 formats for the address in net.sock.peer.addr if net.sock.family is not set. This is to support instrumentations that follow previous versions of this document.

[5]: If different than net.peer.name and if net.sock.peer.addr is set.

[6]: If defined for the address family and if different than net.peer.port and if net.sock.peer.addr is set.

messaging.destination_kind MUST be one of the following:

Additionally net.peer.port from the network attributes is recommended. Furthermore, it is strongly recommended to add the net.transport attribute and follow its guidelines, especially for in-process queueing systems (like Hangfire, for example). These attributes should be set to the broker to which the message is sent/from which it is received.

For message consumers, the following additional attributes may be set:

messaging.operation MUST be one of the following:

The receive span is be used to track the time used for receiving the message(s), whereas the process span(s) track the time for processing the message(s). Note that one or multiple Spans with messaging.operation = process may often be the children of a Span with messaging.operation = receive. The distinction between receiving and processing of messages is not always of particular interest or sometimes hidden away in a framework (see the Message consumption section above) and therefore the attribute can be left out. For batch receiving and processing (see the Batch receiving and Batch processing examples below) in particular, the attribute SHOULD be set. Even though in that case one might think that the processing span’s kind should be INTERNAL, that kind MUST NOT be used. Instead span kind should be set to either CONSUMER or SERVER according to the rules defined above.

Attributes specific to certain messaging systems

RabbitMQ

In RabbitMQ, the destination is defined by an exchange and a routing key. messaging.destination MUST be set to the name of the exchange. This will be an empty string if the default exchange is used.

Apache Kafka

For Apache Kafka, the following additional attributes are defined:

[1]: If the key type is not string, it’s string representation has to be supplied for the attribute. If the key has no unambiguous, canonical string form, don’t include its value.

[2]: If value is true. When missing, the value is assumed to be false.

For Apache Kafka producers, peer.service SHOULD be set to the name of the broker or service the message will be sent to. The service.name of a Consumer’s Resource SHOULD match the peer.service of the Producer, when the message is directly passed to another service. If an intermediary broker is present, service.name and peer.service will not be the same.

Apache RocketMQ

Specific attributes for Apache RocketMQ are defined below.

messaging.rocketmq.message_type MUST be one of the following:

messaging.rocketmq.consumption_model MUST be one of the following:

Examples

Topic with multiple consumers

Given is a process P, that publishes a message to a topic T on messaging system MS, and two processes CA and CB, which both receive the message and process it.

Process P:  | Span Prod1 |
--
Process CA:              | Span CA1 |
--
Process CB:                 | Span CB1 |

Apache Kafka with Quarkus or Spring Boot Example

Given is a process P, that publishes a message to a topic T1 on Apache Kafka. One process, CA, receives the message and publishes a new message to a topic T2 that is then received and processed by CB.

Frameworks such as Quarkus and Spring Boot separate processing of a received message from producing subsequent messages out. For this reason, receiving (Span Rcv1) is the parent of both processing (Span Proc1) and producing a new message (Span Prod2). The span representing message receiving (Span Rcv1) should not set messaging.operation to receive, as it does not only receive the message but also converts the input message to something suitable for the processing operation to consume and creates the output message from the result of processing.

Process P:  | Span Prod1 |
--
Process CA:              | Span Rcv1 |
                                | Span Proc1 |
                                  | Span Prod2 |
--
Process CB:                           | Span Rcv2 |

Batch receiving

Given is a process P, that sends two messages to a queue Q on messaging system MS, and a process C, which receives both of them in one batch (Span Recv1) and processes each message separately (Spans Proc1 and Proc2).

Since a span can only have one parent and the propagated trace and span IDs are not known when the receiving span is started, the receiving span will have no parent and the processing spans are correlated with the producing spans using links.

Process P: | Span Prod1 | Span Prod2 |
--
Process C:                      | Span Recv1 |
                                        | Span Proc1 |
                                               | Span Proc2 |

Batch processing

Given is a process P, that sends two messages to a queue Q on messaging system MS, and a process C, which receives both of them separately (Span Recv1 and Recv2) and processes both messages in one batch (Span Proc1).

Since each span can only have one parent, C3 should not choose a random parent out of C1 and C2, but rather rely on the implicitly selected parent as defined by the tracing API spec. Similarly, only one value can be set as message_id, so C3 cannot report both a1 and a2 and therefore attribute is left out. Depending on the implementation, the producing spans might still be available in the meta data of the messages and should be added to C3 as links. The client library or application could also add the receiver span’s SpanContext to the data structure it returns for each message. In this case, C3 could also add links to the receiver spans C1 and C2.

The status of the batch processing span is selected by the application. Depending on the semantics of the operation. A span status Ok could, for example, be set only if all messages or if just at least one were properly processed.

Process P: | Span Prod1 | Span Prod2 |
--
Process C:                              | Span Recv1 | Span Recv2 |
                                                                   | Span Proc1 |