Trace-level latency widget

[sidhi-ts]

Describe your question

I want to use the latest unstable/dashboard-widgets api to create a trace-level view widget to get latency stats. However, the documentation mentions: The legacy traces view is not supported by this unstable API. I am able to do this from the UI.

Does this mean that the trace level view will never be supported? How can I create a widget for trace level latency using the observation view via SDK?

Langfuse Cloud or Self-Hosted?

Self-Hosted

If Self-Hosted

v3.224.2

If Langfuse Cloud

No response

SDK and integration versions

SDK v.4.14.1

Pre-Submission Checklist

• I have checked for existing issues/discussions and consulted Langfuse AI.

[inkeep-ai-assistant[bot]]

The traces view is excluded from the unstable API and widget builder in Langfuse v4 due to performance concerns. In Langfuse v4's data model, the traces view must aggregate multiple rows per trace to reconstruct trace-level fields, which is expensive for large projects.(1)

Why the Traces View Is Excluded

In Langfuse v4, all data lives in a single denormalized ClickHouse table—the wide observations table. The legacy traces view still exists for backward compatibility but is expensive to query at scale. It is no longer possible to create new custom widgets relying on this view, and existing widgets should be migrated away from it.(1)

How to Get Trace-Level Latency Using Observations

Trace counts and trace-level metrics are now derived from observations. Specifically:

• Trace counts are computed as uniq(trace_id) from the wide observations table—counting distinct trace IDs across all observations(1)
• Trace time bucketing uses the root observation's start_time(1)

To create a trace-level latency widget using the observations view, you would group by trace_id (or name if you want latency by trace name) and use latency metrics. However, note that some very-high-cardinality fields (id, traceId, parentObservationId) are no longer available as dimensions in the custom widget builder because grouping by individual trace ID is not meaningful for charts.(1)

Practical Approach

If you need latency percentiles (p50, p90, p95, p99), you can create widgets using the observations view with appropriate dimensions like name and latency aggregations. The observations-level latency metrics query the wide observations table directly and remain performant.(1)

There is also an open issue regarding trace latency percentile queries timing out on large observations tables, which the team is addressing more fundamentally in the new data model.(2)

📚 Sources:

1. Dashboard Changes in Langfuse v4
2. bug: Trace Latency Percentiles Panel Stuck Loading Due to SQL Timeout on Large observations Tables

---

Have another question? Just tag @inkeep.

[CodeCLS]

No, trace-level widgets aren't ruled out. The restriction sits only on what the create and update endpoints accept, right now that's the observations and score views. Widgets you build in the UI with the trace view still come back fine when you read them via the API. In the meantime you can get trace latency from the observations view. A trace's latency is just its root span's latency, so build a latency widget on observations and filter it down to root spans only. Those are the observations with no parent observation. Group by trace name if you want a breakdown per trace type.