Add the ability to store traces locally instead of sending them to the backend

[XyAgent]

Facilitate Evaluation / BDD and TDD flow

Currently, langfuse provides a very convenient set up to catch application traces, including Agentic (Langgraph-like) traces with detailed steps description, intermediate data transformations, etc.

All those traces and up being stored inside Clickhouse (either launched locally or on a dedicated machine/cloud)
Which is suitable for monitoring, but, sadly, not very convenient for debugging and evaluation purposes

We end up using deepeval, which provides an ability to launch experiments and save evaluation results locally (in a .json file), but I should tell that we are very unhappy to maintain a heavy additional dependency + set up a duplicate observation logic

If only langfuse would provide a way to store traces locally (inside an .sqlite base, for example, or just a plain .json)
this would allow to:

• Having langfuse monitoring setup done (observations, spans, traces logic):
  • Run pytests against test run results, exploring traces and capturing bugs/regressions
  • Set up AI-assistant automatic refinement process (basically TDD) - for example tune prompts until you have a clear run
  • Be free from heavy external Langfuse dependencies (a bunch of services), which is helpful if they are currently unavailable etc.

Making Langfuse an all-in-one solution for testing, evaluation, monitoring, which is currently a pain point

P.S. I thought about langfuse self-hosted runs to address this problem, but:

1. Running PostgreSQL, ClickHouse, Redis/Valkey, S3/Blob storage, Worker container for simple tests looks like a big overhead
2. Exporting traces to self-hosted langfuse and then re-importing them via sdk looks cumbersome

P.S.S.
The thing is not about sqlite or any other particular instrument to store the traces locally, it's more about an abstract class
which would provide an ability to attach a custom trace exporter (which would flush the traces whenever the user wants to)

Just like attaching a custom httpx client when initializing Langfuse object, but for traces

Additional information

No response

[jannikmaierhoefer]

Hi @XyAgent, thanks for adding this.