Before using this dashboard, instrument your Dify applications with OpenTelemetry and configure export to SigNoz. See the Dify observability guide for complete setup instructions.
This dashboard provides a comprehensive view of the langgenius/dify service using trace data. It covers RED metrics (Rate, Errors, Duration), HTTP endpoint performance, LLM/GenAI activity (token usage, model latency, and provider breakdown), Dify app activity, streaming mode distribution, and Postgres/Redis dependency health.
Dashboard Preview
Dashboards → + New dashboard → Import JSON
What This Dashboard Monitors
This dashboard tracks critical performance metrics for your Dify LLM service using OpenTelemetry trace data to help you:
- Measure Traffic and Errors: See the overall request rate (spans/min) and error span count at a glance so you can detect incidents immediately.
- Track End-to-End Latency: Monitor p95 latency on root spans to surface regressions and ensure consistent user-facing responsiveness.
- Optimize LLM Cost: Break down input and output token consumption per model to understand cost drivers and track usage trends.
- Identify Slow Operations: Drill into per-operation request counts and p95 latency to pinpoint the slowest HTTP endpoints.
- Monitor LLM Providers: Compare call volumes and response latency across every AI model in use to guide provider selection.
- Understand App Activity: See which Dify apps and streaming configurations are generating the most traffic.
- Watch Dependencies: Track Postgres query latency and Redis operation counts to catch database bottlenecks early.
Panels Included
RED Metrics (Summary Row)
| Panel | Type | What It Shows |
|---|---|---|
| Request Rate (spans/min) | Stat | Average spans per minute, showing the overall traffic level for the service |
| Error Spans | Stat | Count of spans with has_error = true; turns red when any errors are present |
| p95 Latency (root spans) | Stat | 95th-percentile latency for root spans (parent_span_id = '') |
| Total LLM Tokens | Stat | Sum of gen_ai.usage.total_tokens across all LLM spans in the selected window |
HTTP Operation Breakdown
- Requests by Operation: Table showing span count per operation name to reveal which endpoints receive the most traffic.
- p95 Latency by Operation: Table of p95 duration per operation name to quickly identify the slowest operations.
LLM / GenAI Activity
- LLM p95 Latency by Model: Time-series graph of p95 latency broken down by
gen_ai.request.modelfor comparing response time across providers. - LLM Calls by Model: Bar chart of LLM invocation counts per model over time, showing model adoption and call volume trends.
- Token Usage: Input vs Output (by model): Table of total input and output tokens per model, useful for per-model cost tracking.
Dify App Activity
- Activity by Dify App: Table of request counts grouped by
dify.app_id, showing which Dify applications are most active. - Streaming vs Non-Streaming Requests: Pie chart of the
dify.streamingattribute, showing the proportion of streaming vs. non-streaming traffic.
Dependency Health
- Postgres Query p95 Latency: Line graph of p95 latency for PostgreSQL operations (
db.system = 'postgresql'), grouped bydb.operation, to help catch slow queries. - Redis Operations by Type: Line graph of Redis operation counts (
db.system = 'redis'), grouped by span name, revealing cache and queue activity patterns.
Endpoint Performance Table
- Slowest Endpoints (by p95): Full-width table listing HTTP endpoints filtered by
http.method EXISTS, showing p95 latency and call count side by side for quick performance triage.