Langfuse is open-source and self-hostable, so user inputs, prompts, and judge outputs never leave the customer's VPC, that residency story is what regulated buyers reach for over LangSmith's hosted-only path.
Imagine two security camera companies. Both record what happens in your warehouse and give you a nice video player to review the footage. The first company stores all the recordings in their own data center, and you watch through their website. The second company sells you the camera, the recorder, and the player, and everything lives inside your warehouse's own network. Most small shops happily use the first one because it is easier to set up. But a hospital, a bank, or a government office is required to keep recordings of sensitive activity on their own premises. They pick the second one even though it is more work to install. Langfuse is the second one. LangSmith is the first one.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Picking an LLM observability backend looks like a feature comparison and turns out to be a compliance comparison the moment the workload becomes regulated. For consumer products, internal tools, and most B2B SaaS, LangSmith is an excellent choice and its tight integration with LangChain and LangGraph makes it the path of least resistance. For healthcare, finance, government, and any enterprise with strict residency commitments, the conversation collapses to which vendors can deploy inside the customer's perimeter, and Langfuse is the most common answer at that tier in 2026.
The shape of the argument is worth understanding because it appears repeatedly across the LLMOps stack. Tools that hold prompts, traces, judge scores, datasets, and user inputs are handling the same data the application itself handles. If that data is sensitive in the application, it is sensitive in the observability tool, and the same residency rules apply.
This card walks through what residency actually means for an LLM trace store, how Langfuse's deployment model satisfies it, what self-hosting costs in operational terms, and when the choice flips back to LangSmith or Langfuse Cloud.
What residency means for an LLM trace store
A typical LLM application trace contains the system prompt, the user message, any retrieved documents from RAG, the model's output, any tool-call arguments and results, latency and cost metrics, and possibly a judge score with its reasoning. In a customer-support assistant, the user message often contains account identifiers, names, support tickets, or PHI in healthcare contexts. In a financial workflow, retrieved documents may include account positions, trade details, or regulatory filings. In a government application, the input may be subject to controlled unclassified information rules.
A hosted observability vendor receives all of that data over the wire and stores it in their own infrastructure. The vendor's data-processing terms, security attestations, and physical data-center locations now become part of the customer's compliance picture. For a SOC2 or ISO 27001 customer that wants minimum sub-processors, this adds one. For a HIPAA-covered entity, it requires a BAA with the vendor. For a GDPR data controller, it requires updating the sub-processor list. For a government customer with FedRAMP requirements, it usually requires the vendor to hold the equivalent attestation.
Self-hosting collapses all of that. The trace store is just another internal database, owned by the customer's existing infrastructure team, covered by the existing infrastructure attestation, located in the regions the customer already operates in. No new vendor relationship, no new DPA, no new sub-processor. The compliance burden becomes operational instead of contractual.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- A US healthcare scale-up running Claude Sonnet 4 on Bedrock self-hosts Langfuse inside their AWS account so PHI in user prompts never leaves the same VPC that already holds their patient data.
- A European bank using Gemini 2.5 on Vertex AI deploys Langfuse on GKE inside their existing GCP project to honor GDPR residency commitments without adding a new sub-processor.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you operate Langfuse in production at non-trivial trace volume?
Treat ClickHouse as a primary database, sized for trace ingest, backed up, monitored; tune retention policies aggressively because trace volume grows linearly with traffic.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Picking based on UI polish or LangGraph integration. The decisive factor for regulated buyers is whether the traces leave the perimeter, and that is a deployment-model question.
60 second bullets to scan on the way to the call.
Why residency drives the choice for regulated buyers
Langfuse deployment shape: Postgres, ClickHouse, worker, Next.js UI
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.