Span
The fundamental unit of distributed tracing, representing a single operation or step within a complete trace. It contains metadata such as inputs and outputs, timing, status, and parent-child relationships, enabling precise identification of where issues occur.
Core Facts
In agent scenarios, a complete agent task should be treated as a single Trace, with each LLM call, tool call, and memory retrieval being a Span.
An AI Agent session is essentially an independent runtime context that completes tool registration during the initialization phase and fixes the tool registry as a snapshot from startup time; active sessions cannot dynamically detect configuration changes.
Timeline (last 90 days)
The solution is to raise the sampling unit from Span to Run: a run is either fully recorded or fully summarized, with no partial runs or orphan Spans.
In distributed tracing, a Trace represents one complete end-to-end request flow, a Run usually refers to one complete task lifecycle of Agent execution, and a Span is the smallest unit of work within a Trace; a Run often contains dozens or even hundreds of Spans, forming a tree-shaped causal relationship through parent_id.
Storing high-dimensional embedding vectors (such as 768- or 1536-dimensional float arrays) in span attributes bloats the payload, increases serialization overhead, exhausts the memory of downstream storage engines, and is unreadable to humans.
A span is the smallest unit of work in distributed tracing, representing an operation with a start time, end time, and metadata, and multiple spans are combined into a trace through parent-child relationships or links.
In LLM applications, a single user question typically corresponds to one trace, and the internal model calls, vector retrievals, and tool executions are each a child span.
OTel的核心数据模型是Span——代表一次操作的结构化记录，包含名称、起止时间、状态、属性（attributes）以及与其他span的父子关系；Trace由一组共享同一trace ID的span构成的有向无环图。
An execution trace refers to a structured record of all operational steps throughout a complete task execution, including the inputs and outputs of each LLM call, tool call parameters and return values, intermediate reasoning states, and timestamps with timing information.
The trace structure displayed in Phoenix is essentially a tree of Spans, where each Span represents a discrete operation and records the start time, end time, input parameters, and return values.
Snapshotting key configuration and permission scopes as Span attributes into traces is an effective means of guarding against configuration drift issues.
Mainstream tracing frameworks such as OpenTelemetry generate Spans by instrumenting code and can only record what is explicitly specified to be recorded.
2 more timeline events
All Facts (14)
In agent scenarios, a complete agent task should be treated as a single Trace, with each LLM call, tool call, and memory retrieval being a Span.
65%VerifiedAn AI Agent session is essentially an independent runtime context that completes tool registration during the initialization phase and fixes the tool registry as a snapshot from startup time; active sessions cannot dynamically detect configuration changes.
65%Unverified事件日志中的事件天然对应分布式追踪中的Span概念，每个事件可携带TraceID和SpanID来还原完整的因果链路
85%UnverifiedThe solution is to raise the sampling unit from Span to Run: a run is either fully recorded or fully summarized, with no partial runs or orphan Spans.
50%UnverifiedIn distributed tracing, a Trace represents one complete end-to-end request flow, a Run usually refers to one complete task lifecycle of Agent execution, and a Span is the smallest unit of work within a Trace; a Run often contains dozens or even hundreds of Spans, forming a tree-shaped causal relationship through parent_id.
50%UnverifiedA span is the smallest unit of work in distributed tracing, representing an operation with a start time, end time, and metadata, and multiple spans are combined into a trace through parent-child relationships or links.
50%UnverifiedStoring high-dimensional embedding vectors (such as 768- or 1536-dimensional float arrays) in span attributes bloats the payload, increases serialization overhead, exhausts the memory of downstream storage engines, and is unreadable to humans.
50%UnverifiedIn LLM applications, a single user question typically corresponds to one trace, and the internal model calls, vector retrievals, and tool executions are each a child span.
50%UnverifiedOTel的核心数据模型是Span——代表一次操作的结构化记录，包含名称、起止时间、状态、属性（attributes）以及与其他span的父子关系；Trace由一组共享同一trace ID的span构成的有向无环图。
50%UnverifiedAn execution trace refers to a structured record of all operational steps throughout a complete task execution, including the inputs and outputs of each LLM call, tool call parameters and return values, intermediate reasoning states, and timestamps with timing information.
50%UnverifiedThe trace structure displayed in Phoenix is essentially a tree of Spans, where each Span represents a discrete operation and records the start time, end time, input parameters, and return values.
50%UnverifiedMainstream tracing frameworks such as OpenTelemetry generate Spans by instrumenting code and can only record what is explicitly specified to be recorded.
50%UnverifiedSnapshotting key configuration and permission scopes as Span attributes into traces is an effective means of guarding against configuration drift issues.
50%UnverifiedThe concept of Trace (distributed tracing) is borrowed from the OpenTelemetry standard in the distributed systems field, with the core idea of generating a globally unique Trace ID for a complete request and passing it to each sub-call (Span).
50%