Principle:Langfuse Langfuse Event Validation and Schema Parsing
| Knowledge Sources | |
|---|---|
| Domains | Data Validation, Trace Ingestion |
| Last Updated | 2026-02-14 00:00 GMT |
Overview
Event validation and schema parsing is the practice of defining strict structural contracts for every ingestion event type and enforcing those contracts at the boundary where raw JSON enters the system.
Description
The Langfuse trace ingestion pipeline accepts a wide variety of event types from multiple SDK languages (Python, JavaScript/TypeScript) and from OpenTelemetry-compatible instrumentations. Each event represents a different kind of telemetry: trace creation, span creation/update, generation creation/update, score creation, SDK log messages, and more. Without rigorous schema validation, malformed or incompatible payloads could silently corrupt analytics data, crash downstream processors, or produce misleading dashboards.
The event validation system addresses this by:
- Enumerating all event types: A central
eventTypesconstant maps human-readable event names (e.g.,"trace-create","generation-update") to their string identifiers. This enumeration serves as the single source of truth for what the pipeline can process.
- Defining body schemas per event type: Each event type has a corresponding Zod v4 schema that validates the event body. For example,
TraceBodyvalidates fields likeid,name,sessionId,environment,tags, andmetadata. Observation-related schemas (spans, generations, events) build on a sharedOptionalObservationBodybase and extend it with type-specific fields likeendTime,model,usage, andusageDetails.
- Handling usage format diversity: Token usage can arrive in multiple formats: the legacy Langfuse format (
promptTokens,completionTokens,totalTokens), the modern Langfuse format (input,output,total,unit), or the OpenAI-native format (prompt_tokens,completion_tokens,total_tokenswith optional detail breakdowns). Transform pipelines within the schema normalize all formats into a unified representation.
- Supporting discriminated unions: The top-level ingestion event schema uses a Zod discriminated union on the
typefield. This ensures that the parser selects the correct body schema based on the event type, providing precise error messages when validation fails.
- Separating public and internal validation: A factory function
createIngestionEventSchemaproduces either a "public" or "internal" variant. The public variant enforces that environment names must not start with "langfuse" (reserved prefix), while the internal variant relaxes this restriction for Langfuse-generated events such as prompt experiment traces.
Usage
Apply this principle whenever:
- Adding a new event type to the ingestion pipeline (e.g., a new observation subtype).
- Changing the structure of an existing event body (fields must be added as optional to preserve backward compatibility).
- Integrating data from a new SDK or instrumentation library that may use a different field naming convention.
- Normalizing data formats (e.g., token counts) from heterogeneous sources into a canonical representation.
Theoretical Basis
The validation system is built on the concept of parse, don't validate, where raw data is transformed into well-typed structures at the system boundary, and all downstream code operates on the validated types without further checking.
Event Type Taxonomy
eventTypes = {
TRACE_CREATE: "trace-create" -- Top-level trace
SCORE_CREATE: "score-create" -- Evaluation score
EVENT_CREATE: "event-create" -- Instant observation event
SPAN_CREATE: "span-create" -- Duration-based span
SPAN_UPDATE: "span-update" -- Span update (partial)
GENERATION_CREATE: "generation-create" -- LLM generation
GENERATION_UPDATE: "generation-update" -- Generation update (partial)
AGENT_CREATE: "agent-create" -- Agent observation
TOOL_CREATE: "tool-create" -- Tool call observation
CHAIN_CREATE: "chain-create" -- Chain observation
RETRIEVER_CREATE: "retriever-create" -- Retriever observation
EVALUATOR_CREATE: "evaluator-create" -- Evaluator observation
EMBEDDING_CREATE: "embedding-create" -- Embedding observation
GUARDRAIL_CREATE: "guardrail-create" -- Guardrail observation
SDK_LOG: "sdk-log" -- SDK diagnostic log
DATASET_RUN_ITEM_CREATE: "dataset-run-item-create" -- Dataset run item (internal only)
OBSERVATION_CREATE: "observation-create" -- Legacy observation (deprecated)
OBSERVATION_UPDATE: "observation-update" -- Legacy observation update (deprecated)
}
Schema Inheritance Hierarchy
OptionalObservationBody
|-- CreateEventEvent (adds required id)
| |-- CreateSpanBody (adds endTime)
| | |-- CreateGenerationBody (adds model, modelParameters, usage,
| | usageDetails, costDetails, promptName, promptVersion)
| |
| |-- UpdateSpanBody (adds endTime)
| |-- UpdateGenerationBody (same extensions as CreateGenerationBody)
|
TraceBody (independent schema with id, name, sessionId, userId, environment, tags, etc.)
ScoreBody (discriminated union on dataType: NUMERIC | CATEGORICAL | BOOLEAN | CORRECTION)
SdkLogEvent (log field only)
Usage Normalization Pipeline
Input: Raw usage object from SDK
|
+-- If contains promptTokens/completionTokens/totalTokens (legacy):
| Transform to { input, output, total, unit: "TOKENS" }
|
+-- If contains prompt_tokens/completion_tokens/total_tokens (OpenAI Completion API):
| Extract detail breakdowns (prompt_tokens_details, completion_tokens_details)
| Compute residual base counts after subtracting detail categories
| Transform to { input, output, total, input_*, output_* }
|
+-- If contains input_tokens/output_tokens/total_tokens (OpenAI Response API):
| Same breakdown logic as above
| Transform to { input, output, total, input_*, output_* }
|
+-- Otherwise: Pass through as raw { key: integer } record
|
Output: Validated Usage or UsageDetails object
ID Constraints
All entity IDs are validated with the following constraints:
- Minimum length: 1 character
- Maximum length: 800 characters (to fit within AWS S3 object key limits of 1024 bytes with room for path prefixes)
- Must not contain carriage return characters (
\r)