Trace comparison helps you understand why two executions of the same workflow produced different results, latencies, or costs. This is useful for debugging regressions, A/B testing prompts, and analyzing performance variations.Documentation Index
Fetch the complete documentation index at: https://docs.getnetra.ai/llms.txt
Use this file to discover all available pages before exploring further.
When to Compare Traces
- Debugging regressions - Compare a failing trace with a successful one to spot differences
- A/B testing - Compare traces from different prompt versions or model configurations
- Performance analysis - Identify why some requests are slower than others
- Cost optimization - Compare token usage across different approaches
Accessing Trace Comparison
- Navigate to Observability → Traces in the Netra dashboard
- Select the first trace you want to compare by clicking on it
- Hold
Shiftand click on the second trace, or use the checkbox selection - Click the Compare button in the toolbar
Understanding the Comparison View
The comparison view displays two traces side-by-side, making it easy to spot differences in execution flow, timing, and outcomes.
- Trace ID and timestamp at the top of each panel
- Hierarchical span structure showing the full execution flow (e.g.,
pdf-qa-query→retrieval→query-embedding→openai.embedding) - Duration for each span displayed on the right side
- Color coding — error spans are shown in red, successfully executed spans in green
Comparison Scenarios
Scenario 1: Debugging a Slow Request
You notice some requests taking 5x longer than usual. Compare a slow trace with a fast one:- Filter traces by duration to find a slow example
- Find a similar request with normal duration
- Compare the two traces
- Spans that are significantly slower in one trace
- Additional spans in the slow trace (retry loops, fallbacks)
- Different code paths being executed
Scenario 2: Comparing Prompt Versions
You’ve updated a prompt and want to verify the change:- Run both prompt versions with the same input
- Capture the trace IDs or find them by timestamp
- Compare the traces
- Token count differences (is the new prompt more efficient?)
- Response quality differences in the completion content
- Latency changes
Scenario 3: Investigating Failures
A request that usually succeeds is now failing:- Find a recent failed trace
- Find a recent successful trace for the same operation
- Compare to identify the divergence point
- The span where the error occurred
- Input differences that might have caused the failure
- Missing or different API responses
Tips for Effective Comparison
- Use consistent naming - Ensure your spans have consistent names across traces so they align properly in comparison view.
-
Add identifying attributes - Include attributes like
request.id,user.id, orexperiment.variantto make traces easier to find and compare. - Compare similar operations - Comparing traces from completely different workflows won’t be useful. Filter to the same operation type.
- Check timestamps - Ensure you’re comparing traces from similar time periods to avoid infrastructure differences.
Exporting Comparison Data
To share comparison results with your team:- Click Export in the comparison view
- Choose your format:
- PDF - Visual report with timeline and metrics
- JSON - Raw data for programmatic analysis
- The export includes both traces and highlighted differences
Next Steps
- Custom Trace Views - Create saved views to quickly find traces for comparison
- Traces Overview - Learn more about traces and spans