Tracing API

Status: Stable, Feature-freeze

The Tracing API consist of these main classes:

  • TracerProvider is the entry point of the API. It provides access to Tracers.
  • Tracer is the class responsible for creating Spans.
  • Span is the API to trace an operation.

Data types

While languages and platforms have different ways of representing data, this section defines some generic requirements for this API.

Time

OpenTelemetry can operate on time values up to nanosecond (ns) precision. The representation of those values is language specific.

Timestamp

A timestamp is the time elapsed since the Unix epoch.

  • The minimal precision is milliseconds.
  • The maximal precision is nanoseconds.

Duration

A duration is the elapsed time between two events.

  • The minimal precision is milliseconds.
  • The maximal precision is nanoseconds.

TracerProvider

Tracers can be accessed with a TracerProvider.

In implementations of the API, the TracerProvider is expected to be the stateful object that holds any configuration.

Normally, the TracerProvider is expected to be accessed from a central place. Thus, the API SHOULD provide a way to set/register and access a global default TracerProvider.

Notwithstanding any global TracerProvider, some applications may want to or have to use multiple TracerProvider instances, e.g. to have different configuration (like SpanProcessors) for each (and consequently for the Tracers obtained from them), or because its easier with dependency injection frameworks. Thus, implementations of TracerProvider SHOULD allow creating an arbitrary number of TracerProvider instances.

TracerProvider operations

The TracerProvider MUST provide the following functions:

  • Get a Tracer

Get a Tracer

This API MUST accept the following parameters:

  • name (required): This name SHOULD uniquely identify the instrumentation scope, such as the instrumentation library (e.g. io.opentelemetry.contrib.mongodb), package, module or class name. If an application or library has built-in OpenTelemetry instrumentation, both Instrumented library and Instrumentation library may refer to the same library. In that scenario, the name denotes a module name or component name within that library or application. In case an invalid name (null or empty string) is specified, a working Tracer implementation MUST be returned as a fallback rather than returning null or throwing an exception, its name property SHOULD be set to an empty string, and a message reporting that the specified value is invalid SHOULD be logged. A library, implementing the OpenTelemetry API may also ignore this name and return a default instance for all calls, if it does not support “named” functionality (e.g. an implementation which is not even observability-related). A TracerProvider could also return a no-op Tracer here if application owners configure the SDK to suppress telemetry produced by this library.
  • version (optional): Specifies the version of the instrumentation scope if the scope has a version (e.g. a library version). Example value: 1.0.0.
  • [since 1.4.0] schema_url (optional): Specifies the Schema URL that should be recorded in the emitted telemetry.
  • [since 1.13.0] attributes (optional): Specifies the instrumentation scope attributes to associate with emitted telemetry.

Tracers are identified by name, version, and schema_url fields. When more than one Tracer of the same name, version, and schema_url is created, it is unspecified whether or under which conditions the same or different Tracer instances are returned. It is a user error to create Tracers with different attributes but the same identity.

The term identical applied to Tracers describes instances where all identifying fields are equal. The term distinct applied to Tracers describes instances where at least one identifying field has a different value.

Implementations MUST NOT require users to repeatedly obtain a Tracer again with the same identity to pick up configuration changes. This can be achieved either by allowing to work with an outdated configuration or by ensuring that new configuration applies also to previously returned Tracers.

Note: This could, for example, be implemented by storing any mutable configuration in the TracerProvider and having Tracer implementation objects have a reference to the TracerProvider from which they were obtained. If configuration must be stored per-tracer (such as disabling a certain tracer), the tracer could, for example, do a look-up with its identity in a map in the TracerProvider, or the TracerProvider could maintain a registry of all returned Tracers and actively update their configuration if it changes.

The effect of associating a Schema URL with a Tracer MUST be that the telemetry emitted using the Tracer will be associated with the Schema URL, provided that the emitted data format is capable of representing such association.

Context Interaction

This section defines all operations within the Tracing API that interact with the Context.

The API MUST provide the following functionality to interact with a Context instance:

  • Extract the Span from a Context instance
  • Combine the Span with a Context instance, creating a new Context instance

The functionality listed above is necessary because API users SHOULD NOT have access to the Context Key used by the Tracing API implementation.

If the language has support for implicitly propagated Context (see here), the API SHOULD also provide the following functionality:

  • Get the currently active span from the implicit context. This is equivalent to getting the implicit context, then extracting the Span from the context.
  • Set the currently active span into a new context, and make that the implicit context. This is equivalent to combining the current implicit context’s values with the Span to create a new context, which is then made the current implicit context.

All the above functionalities operate solely on the context API, and they MAY be exposed as either static methods on the trace module, or as static methods on a class inside the trace module. This functionality SHOULD be fully implemented in the API when possible.

Tracer

The tracer is responsible for creating Spans.

Note that Tracers should usually not be responsible for configuration. This should be the responsibility of the TracerProvider instead.

Tracer operations

The Tracer MUST provide functions to:

SpanContext

A SpanContext represents the portion of a Span which must be serialized and propagated along side of a distributed context. SpanContexts are immutable.

The OpenTelemetry SpanContext representation conforms to the W3C TraceContext specification. It contains two identifiers - a TraceId and a SpanId - along with a set of common TraceFlags and system-specific TraceState values.

TraceId A valid trace identifier is a 16-byte array with at least one non-zero byte.

SpanId A valid span identifier is an 8-byte array with at least one non-zero byte.

TraceFlags contain details about the trace. Unlike TraceState values, TraceFlags are present in all traces. The current version of the specification only supports a single flag called sampled.

TraceState carries vendor-specific trace identification data, represented as a list of key-value pairs. TraceState allows multiple tracing systems to participate in the same trace. It is fully described in the W3C Trace Context specification. For specific OTel values in TraceState, see the TraceState Handling document.

The API MUST implement methods to create a SpanContext. These methods SHOULD be the only way to create a SpanContext. This functionality MUST be fully implemented in the API, and SHOULD NOT be overridable.

Retrieving the TraceId and SpanId

The API MUST allow retrieving the TraceId and SpanId in the following forms:

  • Hex - returns the lowercase hex encoded TraceId (result MUST be a 32-hex-character lowercase string) or SpanId (result MUST be a 16-hex-character lowercase string).
  • Binary - returns the binary representation of the TraceId (result MUST be a 16-byte array) or SpanId (result MUST be an 8-byte array).

The API SHOULD NOT expose details about how they are internally stored.

IsValid

An API called IsValid, that returns a boolean value, which is true if the SpanContext has a non-zero TraceID and a non-zero SpanID, MUST be provided.

IsRemote

An API called IsRemote, that returns a boolean value, which is true if the SpanContext was propagated from a remote parent, MUST be provided. When extracting a SpanContext through the Propagators API, IsRemote MUST return true, whereas for the SpanContext of any child spans it MUST return false.

TraceState

TraceState is a part of SpanContext, represented by an immutable list of string key/value pairs and formally defined by the W3C Trace Context specification. Tracing API MUST provide at least the following operations on TraceState:

  • Get value for a given key
  • Add a new key/value pair
  • Update an existing value for a given key
  • Delete a key/value pair

These operations MUST follow the rules described in the W3C Trace Context specification. All mutating operations MUST return a new TraceState with the modifications applied. TraceState MUST at all times be valid according to rules specified in W3C Trace Context specification. Every mutating operations MUST validate input parameters. If invalid value is passed the operation MUST NOT return TraceState containing invalid data and MUST follow the general error handling guidelines.

Please note, since SpanContext is immutable, it is not possible to update SpanContext with a new TraceState. Such changes then make sense only right before SpanContext propagation or telemetry data exporting. In both cases, Propagators and SpanExporters may create a modified TraceState copy before serializing it to the wire.

Span

A Span represents a single operation within a trace. Spans can be nested to form a trace tree. Each trace contains a root span, which typically describes the entire operation and, optionally, one or more sub-spans for its sub-operations.

Spans encapsulate:

The span name concisely identifies the work represented by the Span, for example, an RPC method name, a function name, or the name of a subtask or stage within a larger computation. The span name SHOULD be the most general string that identifies a (statistically) interesting class of Spans, rather than individual Span instances while still being human-readable. That is, “get_user” is a reasonable name, while “get_user/314159”, where “314159” is a user ID, is not a good name due to its high cardinality. Generality SHOULD be prioritized over human-readability.

For example, here are potential span names for an endpoint that gets a hypothetical account information:

Span NameGuidance
getToo general
get_account/42Too specific
get_accountGood, and account_id=42 would make a nice Span attribute
get_account/{accountId}Also good (using the “HTTP route”)

The Span’s start and end timestamps reflect the elapsed real time of the operation.

For example, if a span represents a request-response cycle (e.g. HTTP or an RPC), the span should have a start time that corresponds to the start time of the first sub-operation, and an end time of when the final sub-operation is complete. This includes:

  • receiving the data from the request
  • parsing of the data (e.g. from a binary or json format)
  • any middleware or additional processing logic
  • business logic
  • construction of the response
  • sending of the response

Child spans (or in some cases events) may be created to represent sub-operations which require more detailed observability. Child spans should measure the timing of the respective sub-operation, and may add additional attributes.

A Span’s start time SHOULD be set to the current time on span creation. After the Span is created, it SHOULD be possible to change its name, set its Attributes, add Events, and set the Status. These MUST NOT be changed after the Span’s end time has been set.

Spans are not meant to be used to propagate information within a process. To prevent misuse, implementations SHOULD NOT provide access to a Span’s attributes besides its SpanContext.

Vendors may implement the Span interface to effect vendor-specific logic. However, alternative implementations MUST NOT allow callers to create Spans directly. All Spans MUST be created via a Tracer.

Span Creation

There MUST NOT be any API for creating a Span other than with a Tracer.

In languages with implicit Context propagation, Span creation MUST NOT set the newly created Span as the active Span in the current Context by default, but this functionality MAY be offered additionally as a separate operation.

The API MUST accept the following parameters:

  • The span name. This is a required parameter.

  • The parent Context or an indication that the new Span should be a root Span. The API MAY also have an option for implicitly using the current Context as parent as a default behavior. This API MUST NOT accept a Span or SpanContext as parent, only a full Context.

    The semantic parent of the Span MUST be determined according to the rules described in Determining the Parent Span from a Context.

  • SpanKind, default to SpanKind.Internal if not specified.

  • Attributes. Additionally, these attributes may be used to make a sampling decision as noted in sampling description. An empty collection will be assumed if not specified.

    The API documentation MUST state that adding attributes at span creation is preferred to calling SetAttribute later, as samplers can only consider information already present during span creation.

  • Links - an ordered sequence of Links, see API definition here.

  • Start timestamp, default to current time. This argument SHOULD only be set when span creation time has already passed. If API is called at a moment of a Span logical start, API user MUST NOT explicitly set this argument.

Each span has zero or one parent span and zero or more child spans, which represent causally related operations. A tree of related spans comprises a trace. A span is said to be a root span if it does not have a parent. Each trace includes a single root span, which is the shared ancestor of all other spans in the trace. Implementations MUST provide an option to create a Span as a root span, and MUST generate a new TraceId for each root span created. For a Span with a parent, the TraceId MUST be the same as the parent. Also, the child span MUST inherit all TraceState values of its parent by default.

A Span is said to have a remote parent if it is the child of a Span created in another process. Each propagators’ deserialization must set IsRemote to true on a parent SpanContext so Span creation knows if the parent is remote.

Any span that is created MUST also be ended. This is the responsibility of the user. API implementations MAY leak memory or other resources (including, for example, CPU time for periodic work that iterates all spans) if the user forgot to end the span.

Determining the Parent Span from a Context

When a new Span is created from a Context, the Context may contain a Span representing the currently active instance, and will be used as parent. If there is no Span in the Context, the newly created Span will be a root span.

A SpanContext cannot be set as active in a Context directly, but by wrapping it into a Span. For example, a Propagator performing context extraction may need this.

During Span creation, a user MUST have the ability to record links to other Spans. Linked Spans can be from the same or a different trace – see Links between spans. Links cannot be added after Span creation.

A Link is structurally defined by the following properties:

  • SpanContext of the Span to link to.
  • Zero or more Attributes further describing the link.

The Span creation API MUST provide:

  • An API to record a single Link where the Link properties are passed as arguments. This MAY be called AddLink. This API takes the SpanContext of the Span to link to and optional Attributes, either as individual parameters or as an immutable object encapsulating them, whichever is most appropriate for the language. Implementations MAY ignore links with an invalid SpanContext.

Links SHOULD preserve the order in which they’re set.

Span operations

With the exception of the function to retrieve the Span’s SpanContext and IsRecording, none of the below may be called after the Span is finished.

Get Context

The Span interface MUST provide:

  • An API that returns the SpanContext for the given Span. The returned value may be used even after the Span is finished. The returned value MUST be the same for the entire Span lifetime. This MAY be called GetContext.

IsRecording

Returns true if this Span is recording information like events with the AddEvent operation, attributes using SetAttributes, status with SetStatus, etc.

After a Span is ended, it usually becomes non-recording and thus IsRecording SHOULD consequently return false for ended Spans. Note: Streaming implementations, where it is not known if a span is ended, are one expected case where IsRecording cannot change after ending a Span.

IsRecording SHOULD NOT take any parameters.

This flag SHOULD be used to avoid expensive computations of a Span attributes or events in case when a Span is definitely not recorded. Note that any child span’s recording is determined independently from the value of this flag (typically based on the sampled flag of a TraceFlags on SpanContext).

This flag may be true despite the entire trace being sampled out. This allows to record and process information about the individual Span without sending it to the backend. An example of this scenario may be recording and processing of all incoming requests for the processing and building of SLA/SLO latency charts while sending only a subset - sampled spans - to the backend. See also the sampling section of SDK design.

Users of the API should only access the IsRecording property when instrumenting code and never access SampledFlag unless used in context propagators.

Set Attributes

A Span MUST have the ability to set Attributes associated with it.

The Span interface MUST provide:

  • An API to set a single Attribute where the attribute properties are passed as arguments. This MAY be called SetAttribute. To avoid extra allocations some implementations may offer a separate API for each of the possible value types.

The Span interface MAY provide:

  • An API to set multiple Attributes at once, where the Attributes are passed in a single method call.

Setting an attribute with the same key as an existing attribute SHOULD overwrite the existing attribute’s value.

Note that the OpenTelemetry project documents certain “standard attributes” that have prescribed semantic meanings.

Note that Samplers can only consider information already present during span creation. Any changes done later, including new or changed attributes, cannot change their decisions.

Add Events

A Span MUST have the ability to add events. Events have a time associated with the moment when they are added to the Span.

An Event is structurally defined by the following properties:

  • Name of the event.
  • A timestamp for the event. Either the time at which the event was added or a custom timestamp provided by the user.
  • Zero or more Attributes further describing the event.

The Span interface MUST provide:

  • An API to record a single Event where the Event properties are passed as arguments. This MAY be called AddEvent. This API takes the name of the event, optional Attributes and an optional Timestamp which can be used to specify the time at which the event occurred, either as individual parameters or as an immutable object encapsulating them, whichever is most appropriate for the language. If no custom timestamp is provided by the user, the implementation automatically sets the time at which this API is called on the event.

Events SHOULD preserve the order in which they are recorded. This will typically match the ordering of the events’ timestamps, but events may be recorded out-of-order using custom timestamps.

Consumers should be aware that an event’s timestamp might be before the start or after the end of the span if custom timestamps were provided by the user for the event or when starting or ending the span. The specification does not require any normalization if provided timestamps are out of range.

Note that the OpenTelemetry project documents certain “standard event names and keys” which have prescribed semantic meanings.

Note that RecordException is a specialized variant of AddEvent for recording exception events.

Set Status

Sets the Status of the Span. If used, this will override the default Span status, which is Unset.

Status is structurally defined by the following properties:

  • StatusCode, one of the values listed below.
  • Optional Description that provides a descriptive message of the Status. Description MUST only be used with the Error StatusCode value. An empty Description is equivalent with a not present one.

Note: The OTLP protocol definition refers to the Description property as message.

StatusCode is one of the following values:

  • Unset
    • The default status.
  • Ok
    • The operation has been validated by an Application developer or Operator to have completed successfully.
  • Error
    • The operation contains an error.

These values form a total order: Ok > Error > Unset. This means that setting Status with StatusCode=Ok will override any prior or future attempts to set span Status with StatusCode=Error or StatusCode=Unset. See below for more specific rules.

The Span interface MUST provide:

  • An API to set the Status. This SHOULD be called SetStatus. This API takes the StatusCode, and an optional Description, either as individual parameters or as an immutable object encapsulating them, whichever is most appropriate for the language. Description MUST be IGNORED for StatusCode Ok & Unset values.

The status code SHOULD remain unset, except for the following circumstances:

An attempt to set value Unset SHOULD be ignored.

When the status is set to Error by Instrumentation Libraries, the Description SHOULD be documented and predictable. The status code should only be set to Error according to the rules defined within the semantic conventions. For operations not covered by the semantic conventions, Instrumentation Libraries SHOULD publish their own conventions, including possible values of Description and what they mean.

Generally, Instrumentation Libraries SHOULD NOT set the status code to Ok, unless explicitly configured to do so. Instrumentation Libraries SHOULD leave the status code as Unset unless there is an error, as described above.

Application developers and Operators may set the status code to Ok.

When span status is set to Ok it SHOULD be considered final and any further attempts to change it SHOULD be ignored.

Analysis tools SHOULD respond to an Ok status by suppressing any errors they would otherwise generate. For example, to suppress noisy errors such as 404s.

Only the value of the last call will be recorded, and implementations are free to ignore previous calls.

UpdateName

Updates the Span name. Upon this update, any sampling behavior based on Span name will depend on the implementation.

Note that Samplers can only consider information already present during span creation. Any changes done later, including updated span name, cannot change their decisions.

Alternatives for the name update may be late Span creation, when Span is started with the explicit timestamp from the past at the moment where the final Span name is known, or reporting a Span with the desired name as a child Span.

Required parameters:

  • The new span name, which supersedes whatever was passed in when the Span was started

End

Signals that the operation described by this span has now (or at the time optionally specified) ended.

Implementations SHOULD ignore all subsequent calls to End and any other Span methods, i.e. the Span becomes non-recording by being ended (there might be exceptions when Tracer is streaming events and has no mutable state associated with the Span).

Language SIGs MAY provide methods other than End in the API that also end the span to support language-specific features like with statements in Python. However, all API implementations of such methods MUST internally call the End method and be documented to do so.

End MUST NOT have any effects on child spans. Those may still be running and can be ended later.

End MUST NOT inactivate the Span in any Context it is active in. It MUST still be possible to use an ended span as parent via a Context it is contained in. Also, any mechanisms for putting the Span into a Context MUST still work after the Span was ended.

Parameters:

  • (Optional) Timestamp to explicitly set the end timestamp. If omitted, this MUST be treated equivalent to passing the current time.

Expect this operation to be called in the “hot path” of production applications. It needs to be designed to complete fast, if not immediately. This operation itself MUST NOT perform blocking I/O on the calling thread. Any locking used needs be minimized and SHOULD be removed entirely if possible. Some downstream SpanProcessors and subsequent SpanExporters called from this operation may be used for testing, proof-of-concept ideas, or debugging and may not be designed for production use themselves. They are not in the scope of this requirement and recommendation.

Record Exception

To facilitate recording an exception languages SHOULD provide a RecordException method if the language uses exceptions. This is a specialized variant of AddEvent, so for anything not specified here, the same requirements as for AddEvent apply.

The signature of the method is to be determined by each language and can be overloaded as appropriate. The method MUST record an exception as an Event with the conventions outlined in the exception semantic conventions document. The minimum required argument SHOULD be no more than only an exception object.

If RecordException is provided, the method MUST accept an optional parameter to provide any additional event attributes (this SHOULD be done in the same way as for the AddEvent method). If attributes with the same name would be generated by the method already, the additional attributes take precedence.

Note: RecordException may be seen as a variant of AddEvent with additional exception-specific parameters and all other parameters being optional (because they have defaults from the exception semantic convention).

Span lifetime

Span lifetime represents the process of recording the start and the end timestamps to the Span object:

  • The start time is recorded when the Span is created.
  • The end time needs to be recorded when the operation is ended.

Start and end time as well as Event’s timestamps MUST be recorded at a time of a calling of corresponding API.

Wrapping a SpanContext in a Span

The API MUST provide an operation for wrapping a SpanContext with an object implementing the Span interface. This is done in order to expose a SpanContext as a Span in operations such as in-process Span propagation.

If a new type is required for supporting this operation, it SHOULD NOT be exposed publicly if possible (e.g. by only exposing a function that returns something with the Span interface type). If a new type is required to be publicly exposed, it SHOULD be named NonRecordingSpan.

The behavior is defined as follows:

  • GetContext MUST return the wrapped SpanContext.
  • IsRecording MUST return false to signal that events, attributes and other elements are not being recorded, i.e. they are being dropped.

The remaining functionality of Span MUST be defined as no-op operations. Note: This includes End, so as an exception from the general rule, it is not required (or even helpful) to end such a Span.

This functionality MUST be fully implemented in the API, and SHOULD NOT be overridable.

SpanKind

SpanKind describes the relationship between the Span, its parents, and its children in a Trace. SpanKind describes two independent properties that benefit tracing systems during analysis.

The first property described by SpanKind reflects whether the Span is a “logical” remote child or parent. By “logical”, we mean that the span is logically a remote child or parent, from the point of view of the library that is being instrumented. Spans with a remote parent are interesting because they are sources of external load. Spans with a remote child are interesting because they reflect a non-local system dependency.

The second property described by SpanKind reflects whether a child Span represents a synchronous call. When a child span is synchronous, the parent is expected to wait for it to complete under ordinary circumstances. It can be useful for tracing systems to know this property, since synchronous Spans may contribute to the overall trace latency. Asynchronous scenarios can be remote or local.

In order for SpanKind to be meaningful, callers SHOULD arrange that a single Span does not serve more than one purpose. For example, a server-side span SHOULD NOT be used directly as the parent of another remote span. As a simple guideline, instrumentation should create a new Span prior to extracting and serializing the SpanContext for a remote call.

Note: there are complex scenarios where a CLIENT span may have a child that is also logically a CLIENT span, or a PRODUCER span might have a local child that is a CLIENT span, depending on how the various libraries that are providing the functionality are built and instrumented. These scenarios, when they occur, should be detailed in the semantic conventions appropriate to the relevant libraries.

These are the possible SpanKinds:

  • SERVER Indicates that the span covers server-side handling of a synchronous RPC or other remote request. This span is often the child of a remote CLIENT span that was expected to wait for a response.
  • CLIENT Indicates that the span describes a request to some remote service. This span is usually the parent of a remote SERVER span and does not end until the response is received.
  • PRODUCER Indicates that the span describes the initiators of an asynchronous request. This parent span will often end before the corresponding child CONSUMER span, possibly even before the child span starts. In messaging scenarios with batching, tracing individual messages requires a new PRODUCER span per message to be created.
  • CONSUMER Indicates that the span describes a child of an asynchronous PRODUCER request.
  • INTERNAL Default value. Indicates that the span represents an internal operation within an application, as opposed to an operations with remote parents or children.

To summarize the interpretation of these kinds:

SpanKindSynchronousAsynchronousRemote IncomingRemote Outgoing
CLIENTyesyes
SERVERyesyes
PRODUCERyesmaybe
CONSUMERyesmaybe
INTERNAL

Concurrency

For languages which support concurrent execution the Tracing APIs provide specific guarantees and safeties. Not all of API functions are safe to be called concurrently.

TracerProvider - all methods are safe to be called concurrently.

Tracer - all methods are safe to be called concurrently.

Span - All methods of Span are safe to be called concurrently.

Event - Events are immutable and safe to be used concurrently.

Link - Links are immutable and safe to be used concurrently.

Included Propagators

See Propagators Distribution for how propagators are to be distributed.

Behavior of the API in the absence of an installed SDK

In general, in the absence of an installed SDK, the Trace API is a “no-op” API. This means that operations on a Tracer, or on Spans, should have no side effects and do nothing. However, there is one important exception to this general rule, and that is related to propagation of a SpanContext: The API MUST return a non-recording Span with the SpanContext in the parent Context (whether explicitly given or implicit current). If the Span in the parent Context is already non-recording, it SHOULD be returned directly without instantiating a new Span. If the parent Context contains no Span, an empty non-recording Span MUST be returned instead (i.e., having a SpanContext with all-zero Span and Trace IDs, empty Tracestate, and unsampled TraceFlags). This means that a SpanContext that has been provided by a configured Propagator will be propagated through to any child span and ultimately also Inject, but that no new SpanContexts will be created.