Made the switch six months ago. Langfuse was fine for early experiments, but scaling up? Not so much.
The main gripes:
* **Vendor lock-in, but self-hosted?** The whole "open-core" promise felt like a trap. You're still tied to their schema, their dashboard, their pace. Want to pipe data somewhere else? Good luck without a major engineering project.
* **Cost got weird fast.** Not the SaaS price, but the hidden infra cost of self-hosting. Once you're ingesting serious volume, you're managing Postgres, Redis, object storage... and the bill from your own cloud account starts to look a lot like a vendor invoice.
* **Datadog just works.** Their LLM Observability is just another facet of our existing APM, logs, infra metrics. One query to correlate a latency spike in our LLM calls with a specific pod's CPU throttling and the cost impact. Can't beat that.
If you're a tiny startup with one model, Langfuse is cute. For anyone trying to run this in production with real traffic and a need to connect it to the rest of your system? You're just building your own observability platform, poorly.
fight me