How to show all spans (or events) of the given name?

[hnykda]

Hi.

I only seem to be able to see either traces on /traces or generations /generations. But in none of them I am able to list e.g. all spans and filter by name to e.g. see their latency etc. Is that not possible?

[dosubot[bot]]

To show all spans or events of a given name and filter them to see details like latency, you can use the LatencyTables component. This component queries and displays latency data for traces, generations, and spans. Here's a brief overview of how it works:

1. Querying Data: The component uses api.dashboard.chart.useQuery to fetch data from the traces_observations table. It selects duration percentiles (50th, 90th, 95th, 99th) and groups the data by the name column for spans.

2. Filtering: The queries apply filters based on the globalFilterState and specific type filters (e.g., type: "SPAN" for spans).

3. Displaying Data: The DashboardTable component is used to display the data, showing the name of the span along with the latency percentiles.

This setup allows you to view and filter spans by their name and see detailed latency information ^[1].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}

[marcklingen]

Currently, this is not possible yet. Adding a separate table for all spans has been on the roadmap for some time already.

[hnykda]

Cool, thanks. I am very grateful for langfuse!

Just in case it's not clear, I see this as a relatively big limitation (I am used to it from other tracing solutions), making it borderline unusable for my usecase (I guess I can recreate that using API and so on, but that's not what I really hope for 😁 ).

[marcklingen]

Thanks for sharing! Can you share more details on what you are trying to achieve here? Would love to learn more

[hnykda]

Sure, sure.

One of the main things is that as the LLM applications are more and more complex, prompts stop being useful as the right unit of abstraction for analyzing what your app does. So, single generation doesn't cut it, as the logical unit we work on is e.g. "chain dealing with googling, selecting what to read, extracting, parsing, distilling", or "agent finding a dataset X" with various tools. During one of these chains it can be many LLM generations and explicit custom python/TS/* tooling, and we want optimize these tools as one chunk. I would naturally call these observations for a single run of this tool as a span. During one trace - e.g. user interaction - there are many such chains/spans happening.

I want to be able to display "all spans of type X" to see how expensive a specific type of span was, or input/outputs of all of these spans so I can e.g. export them into a dataset or inputs/outputs for evals etc.

Makes sense? Is there a better way to achieve this with langfuse?

[hnykda]

[hnykda]

Oh, another usecase that I would want, and that is something like how many spans with name X have been executed within this trace.