Your LLM traces are PII, which is the case for self-hosting Langfuse

LLM observability has a property normal APM doesn't: the payload is the user data. Traces contain full prompts and completions — names, emails, medical questions, source code. Shipping them to a third-party SaaS is a data-processing decision, not a tooling decision.

That's the actual argument for running Langfuse yourself rather than the hosted version, and it's stronger than the cost argument. Self-hosted Langfuse is a docker-compose stack; the observability data never leaves your infrastructure, which shortens a lot of compliance conversations to "it's on our box."

The trade is operational: you own Postgres and ClickHouse uptime, and traces are the first thing you'll want during an incident, which is precisely when your own infra is least trustworthy. For a solo project that risk is fine; for a team, budget for it.

Setup and what the hosted free tier covers if you go the other way: https://toolfreebie.com/langfuse-llm-observability/