General RPC conventions

Status: Experimental

The conventions described in this section are RPC specific. When RPC operations occur, metric events about those operations will be generated and reported to provide insight into those operations. By adding RPC properties as attributes on metric events it allows for finely tuned filtering.

Metric instruments

The following metric instruments MUST be used to describe RPC operations. They MUST be of the specified type and units.

Note: RPC server and client metrics are split to allow correlation across client/server boundaries, e.g. Lining up an RPC method latency to determine if the server is responsible for latency the client is seeing.

RPC Server

Below is a table of RPC server metric instruments.

NameInstrument Type (*)UnitUnit (UCUM)DescriptionStatusStreaming
rpc.server.durationHistogrammillisecondsmsmeasures duration of inbound RPCRecommendedN/A. While streaming RPCs may record this metric as start-of-batch to end-of-batch, it’s hard to interpret in practice.
rpc.server.request.sizeHistogramBytesBymeasures size of RPC request messages (uncompressed)OptionalRecorded per message in a streaming batch
rpc.server.response.sizeHistogramBytesBymeasures size of RPC response messages (uncompressed)OptionalRecorded per response in a streaming batch
rpc.server.requests_per_rpcHistogramcount{count}measures the number of messages received per RPC. Should be 1 for all non-streaming RPCsOptionalRequired
rpc.server.responses_per_rpcHistogramcount{count}measures the number of messages sent per RPC. Should be 1 for all non-streaming RPCsOptionalRequired

RPC Client

Below is a table of RPC client metric instruments. These apply to traditional RPC usage, not streaming RPCs.

NameInstrument Type (*)UnitUnit (UCUM)DescriptionStatusStreaming
rpc.client.durationHistogrammillisecondsmsmeasures duration of outbound RPCRecommendedN/A. While streaming RPCs may record this metric as start-of-batch to end-of-batch, it’s hard to interpret in practice.
rpc.client.request.sizeHistogramBytesBymeasures size of RPC request messages (uncompressed)OptionalRecorded per message in a streaming batch
rpc.client.response.sizeHistogramBytesBymeasures size of RPC response messages (uncompressed)OptionalRecorded per message in a streaming batch
rpc.client.requests_per_rpcHistogramcount{count}measures the number of messages received per RPC. Should be 1 for all non-streaming RPCsOptionalRequired
rpc.client.responses_per_rpcHistogramcount{count}measures the number of messages sent per RPC. Should be 1 for all non-streaming RPCsOptionalRequired

Attributes

Below is a table of attributes that SHOULD be included on metric events and whether or not they should be on the server, client or both.

AttributeTypeDescriptionExamplesRequired
rpc.systemstringA string identifying the remoting system. See below for a list of well-known identifiers.grpcYes
rpc.servicestringThe full (logical) name of the service being called, including its package name, if applicable. [1]myservice.EchoServiceNo, but recommended
rpc.methodstringThe name of the (logical) method being called, must be equal to the $method part in the span name. [2]exampleMethodNo, but recommended
net.peer.ipstringRemote address of the peer (dotted decimal for IPv4 or RFC5952 for IPv6)127.0.0.1See below
net.peer.namestringRemote hostname or similar, see note below. [3]example.comSee below
net.peer.portintRemote port number.80; 8080; 443See below
net.transportstringTransport protocol used. See note below.ip_tcpSee below

[1]: This is the logical name of the service from the RPC interface perspective, which can be different from the name of any implementing class. The code.namespace attribute may be used to store the latter (despite the attribute name, it may include a class name; e.g., class with method actually executing the call on the server side, RPC client stub class on the client side).

[2]: This is the logical name of the method from the RPC interface perspective, which can be different from the name of any implementing method/function. The code.function attribute may be used to store the latter (e.g., method actually executing the call on the server side, RPC client stub method on the client side).

[3]: net.peer.name SHOULD NOT be set if capturing it would require an extra DNS lookup.

Additional attribute requirements: At least one of the following sets of attributes is required:

rpc.system has the following list of well-known values. If one of them applies, then the respective value MUST be used, otherwise a custom value MAY be used.

ValueDescription
grpcgRPC
java_rmiJava RMI
dotnet_wcf.NET WCF
apache_dubboApache Dubbo

To avoid high cardinality, implementations should prefer the most stable of net.peer.name or net.peer.ip, depending on expected deployment profile. For many cloud applications, this is likely net.peer.name as names can be recycled even across re-instantiation of a server with a different ip.

For client-side metrics net.peer.port is required if the connection is IP-based and the port is available (it describes the server port they are connecting to). For server-side spans net.peer.port is optional (it describes the port the client is connecting from). Furthermore, setting net.transport is required for non-IP connection like named pipe bindings.

Service name

On the server process receiving and handling the remote procedure call, the service name provided in rpc.service does not necessarily have to match the service.name resource attribute. One process can expose multiple RPC endpoints and thus have multiple RPC service names. From a deployment perspective, as expressed by the service.* resource attributes, it will be treated as one deployed service with one service.name.

gRPC conventions

For remote procedure calls via gRPC, additional conventions are described in this section.

rpc.system MUST be set to "grpc".