Merge branch '1.12.1-otel-ee' into deploy/enterprise

This commit is contained in:
GareArc
2026-02-28 17:41:17 -08:00
4 changed files with 453 additions and 758 deletions

View File

@@ -945,6 +945,17 @@ class TraceTask:
"embedding_model_provider": row[2] or "",
}
# Extract rerank model info from retrieval_model kwargs
rerank_model_provider = ""
rerank_model_name = ""
if "retrieval_model" in kwargs:
retrieval_model = kwargs["retrieval_model"]
if isinstance(retrieval_model, dict):
reranking_model = retrieval_model.get("reranking_model")
if isinstance(reranking_model, dict):
rerank_model_provider = reranking_model.get("reranking_provider_name", "")
rerank_model_name = reranking_model.get("reranking_model_name", "")
metadata = {
"message_id": message_id,
"ls_provider": message_data.model_provider,
@@ -961,6 +972,8 @@ class TraceTask:
"app_name": app_name,
"workspace_name": workspace_name,
"embedding_models": embedding_models,
"rerank_model_provider": rerank_model_provider,
"rerank_model_name": rerank_model_name,
}
if node_execution_id := kwargs.get("node_execution_id"):
metadata["node_execution_id"] = node_execution_id

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,116 @@
# Dify Enterprise Telemetry
This document provides an overview of the Dify Enterprise OpenTelemetry (OTEL) exporter and how to configure it for integration with observability stacks like Prometheus, Grafana, Jaeger, or Honeycomb.
## Overview
Dify Enterprise uses a "slim span + rich companion log" architecture to provide high-fidelity observability without overwhelming trace storage.
- **Traces (Spans)**: Capture the structure, identity, and timing of high-level operations (Workflows and Nodes).
- **Structured Logs**: Provide deep context (inputs, outputs, metadata) for every event, correlated to spans via `trace_id` and `span_id`.
- **Metrics**: Provide 100% accurate counters and histograms for usage, performance, and error tracking.
### Signal Architecture
```mermaid
graph TD
A[Workflow Run] -->|Span| B(dify.workflow.run)
A -->|Log| C(dify.workflow.run detail)
B ---|trace_id| C
D[Node Execution] -->|Span| E(dify.node.execution)
D -->|Log| F(dify.node.execution detail)
E ---|span_id| F
G[Message/Tool/etc] -->|Log| H(dify.* event)
G -->|Metric| I(dify.* counter/histogram)
```
## Configuration
The Enterprise OTEL exporter is configured via environment variables.
| Variable | Description | Default |
|----------|-------------|---------|
| `ENTERPRISE_ENABLED` | Master switch for all enterprise features. | `false` |
| `ENTERPRISE_TELEMETRY_ENABLED` | Master switch for enterprise telemetry. | `false` |
| `ENTERPRISE_OTLP_ENDPOINT` | OTLP collector endpoint (e.g., `http://otel-collector:4318`). | - |
| `ENTERPRISE_OTLP_HEADERS` | Custom headers for OTLP requests (e.g., `x-scope-orgid=tenant1`). | - |
| `ENTERPRISE_OTLP_PROTOCOL` | OTLP transport protocol (`http` or `grpc`). | `http` |
| `ENTERPRISE_OTLP_API_KEY` | Bearer token for authentication. | - |
| `ENTERPRISE_INCLUDE_CONTENT` | Whether to include sensitive content (inputs/outputs) in logs. | `true` |
| `ENTERPRISE_SERVICE_NAME` | Service name reported to OTEL. | `dify` |
| `ENTERPRISE_OTEL_SAMPLING_RATE` | Sampling rate for traces (0.0 to 1.0). Metrics are always 100%. | `1.0` |
## Correlation Model
Dify uses deterministic ID generation to ensure signals are correlated across different services and asynchronous tasks.
### ID Generation Rules
- `trace_id`: Derived from the correlation ID (workflow_run_id or node_execution_id for drafts) using `int(UUID(correlation_id))`
- `span_id`: Derived from the source ID using `SHA256(source_id)[:8]`
### Scenario A: Simple Workflow
A single workflow run with multiple nodes. All spans and logs share the same `trace_id` (derived from `workflow_run_id`).
```
trace_id = UUID(workflow_run_id)
├── [root span] dify.workflow.run (span_id = hash(workflow_run_id))
│ ├── [child] dify.node.execution - "Start" (span_id = hash(node_exec_id_1))
│ ├── [child] dify.node.execution - "LLM" (span_id = hash(node_exec_id_2))
│ └── [child] dify.node.execution - "End" (span_id = hash(node_exec_id_3))
```
### Scenario B: Nested Sub-Workflow
A workflow calling another workflow via a Tool or Sub-workflow node. The child workflow's spans are linked to the parent via `parent_span_id`. Both workflows share the same trace_id.
```
trace_id = UUID(outer_workflow_run_id) ← shared across both workflows
├── [root] dify.workflow.run (outer) (span_id = hash(outer_workflow_run_id))
│ ├── dify.node.execution - "Start Node"
│ ├── dify.node.execution - "Tool Node" (triggers sub-workflow)
│ │ └── [child] dify.workflow.run (inner) (span_id = hash(inner_workflow_run_id))
│ │ ├── dify.node.execution - "Inner Start"
│ │ └── dify.node.execution - "Inner End"
│ └── dify.node.execution - "End Node"
```
**Key attributes for nested workflows:**
- Inner workflow's `dify.parent.trace_id` = outer `workflow_run_id`
- Inner workflow's `dify.parent.node.execution_id` = tool node's `execution_id`
- Inner workflow's `dify.parent.workflow.run_id` = outer `workflow_run_id`
- Inner workflow's `dify.parent.app.id` = outer `app_id`
### Scenario C: Draft Node Execution
A single node run in isolation (debugger/preview mode). It creates its own trace where the node span is the root.
```
trace_id = UUID(node_execution_id) ← own trace, NOT part of any workflow
└── dify.node.execution.draft (span_id = hash(node_execution_id))
```
**Key difference:** Draft executions use `node_execution_id` as the correlation_id, so they are NOT children of any workflow trace.
## Content Gating
When `ENTERPRISE_INCLUDE_CONTENT` is set to `false`, sensitive content attributes (inputs, outputs, queries) are replaced with reference strings (e.g., `ref:workflow_run_id=...`) to prevent data leakage to the OTEL collector.
**Reference String Format:**
```
ref:{id_type}={uuid}
```
**Examples:**
```
ref:workflow_run_id=550e8400-e29b-41d4-a716-446655440000
ref:node_execution_id=660e8400-e29b-41d4-a716-446655440001
ref:message_id=770e8400-e29b-41d4-a716-446655440002
```
To retrieve actual content when gating is enabled, query the Dify database using the provided UUID.
## Reference
For a complete list of telemetry signals, attributes, and data structures, see [DATA_DICTIONARY.md](./DATA_DICTIONARY.md).

View File

@@ -419,9 +419,11 @@ class EnterpriseOtelTrace:
**labels,
type=request_type,
status=info.status,
model_name=info.model_name or "",
),
)
duration_labels = dict(labels)
duration_labels["model_name"] = info.model_name or ""
plugin_name = metadata.get("plugin_name")
if plugin_name and info.node_type in {"tool", "knowledge-retrieval"}:
duration_labels["plugin_name"] = plugin_name
@@ -434,6 +436,7 @@ class EnterpriseOtelTrace:
self._labels(
**labels,
type=request_type,
model_name=info.model_name or "",
),
)
@@ -674,6 +677,8 @@ class EnterpriseOtelTrace:
self._labels(
**labels,
type="suggested_question",
model_provider=info.model_provider or "",
model_name=info.model_id or "",
),
)
@@ -738,6 +743,13 @@ class EnterpriseOtelTrace:
attrs["dify.dataset.embedding_providers"] = self._maybe_json(providers)
attrs["dify.dataset.embedding_models"] = self._maybe_json(models)
# Add rerank model to logs
rerank_provider = metadata.get("rerank_model_provider", "")
rerank_model = metadata.get("rerank_model_name", "")
if rerank_provider or rerank_model:
attrs["dify.retrieval.rerank_provider"] = rerank_provider
attrs["dify.retrieval.rerank_model"] = rerank_model
ref = f"ref:message_id={info.message_id}"
retrieval_inputs = self._safe_payload_value(info.inputs)
attrs["dify.retrieval.query"] = self._content_or_ref(retrieval_inputs, ref)
@@ -766,12 +778,25 @@ class EnterpriseOtelTrace:
)
for did in dataset_ids:
# Get embedding model for this specific dataset
ds_embedding_info = embedding_models.get(did, {})
embedding_provider = ds_embedding_info.get("embedding_model_provider", "")
embedding_model = ds_embedding_info.get("embedding_model", "")
# Get rerank model (same for all datasets in this retrieval)
rerank_provider = metadata.get("rerank_model_provider", "")
rerank_model = metadata.get("rerank_model_name", "")
self._exporter.increment_counter(
EnterpriseTelemetryCounter.DATASET_RETRIEVALS,
1,
self._labels(
**labels,
dataset_id=did,
embedding_model_provider=embedding_provider,
embedding_model=embedding_model,
rerank_model_provider=rerank_provider,
rerank_model=rerank_model,
),
)