Deploy Langfuse v4

Deploy and Host Langfuse on Railway

Langfuse is the open-source LLM engineering platform: tracing, evaluations, prompt management, cost metrics and datasets in one MIT-licensed app. Teams shipping AI agents and RAG systems use it to see which prompt version produced an answer and what a chain cost in tokens. It is OpenTelemetry-native, so instrumenting an app is a decorator, not a rewrite, and tracing needs no LLM provider key.

Self-host Langfuse on Railway behind one public domain. langfuse-web serves the UI and OTLP endpoint; the private langfuse-worker (4.3.1, port 3030) drains the queue. Managed Postgres/Redis, a pinned ClickHouse with a volume and a Railway bucket hold accounts, traces, queues and events.

Getting Started with Langfuse on Railway

Open the generated Railway domain once every service is green. Langfuse seeds no admin account, so the first account created becomes the owner — create it, then set AUTH_DISABLE_SIGNUP=true to close registration. Create an organization and a project, then open Project Settings → API Keys for a key pair; the .env snippet there should carry https://${{RAILWAY_PUBLIC_DOMAIN}}, and localhost means NEXTAUTH_URL is wrong. Send one trace and check the Tracing table — a trace landing there proves the whole path: web → Redis → bucket → worker → ClickHouse.

Sending Your First Trace to Self-Hosted Langfuse

Set LANGFUSE_HOST to https://${{RAILWAY_PUBLIC_DOMAIN}} plus your key pair; the Python OpenAI wrapper is an import swap:

from langfuse.openai import openai

openai.chat.completions.create(model="gpt-4o-mini",
    messages=[{"role": "user", "content": "hi"}])

About Hosting Langfuse

Self-host when prompts cannot sit in a vendor account, or when volume makes per-unit cloud pricing the wrong shape.

• OpenTelemetry tracing — nested traces over LLM calls, retrievals, tool calls and agent steps.
• Prompt management — version and roll back prompts, fetched at runtime via a cached client.
• Evaluations — LLM-as-a-judge, code evaluators, annotation queues.
• SDKs — Python and JS/TS plus LangChain, LlamaIndex, the Vercel AI SDK and LiteLLM.

Why Deploy Langfuse on Railway

Railway runs the whole five-service stack as one project:

• Pinned upstream images — no latest surprises on redeploy.
• ClickHouse with a volume and console logging preset.
• Managed object storage replaces MinIO — one fewer container.
• Health checks, restart policies and a Node heap ceiling preset.

Common Use Cases

• Debugging AI agents — replay a failing run and find the tool call or prompt version behind it.
• Controlling LLM spend — attribute token cost to a user, feature or customer.
• Regression-testing prompts — score each prompt or model change against real traces before rollout.

Dependencies for Langfuse

• langfuse-web — langfuse/langfuse, port 3000, health check /api/public/health. UI, API and OTLP endpoint; runs migrations at boot.
• langfuse-worker — langfuse/langfuse, port 3030, health check /api/health, private. Keep its tag identical to the web image.
• clickhouse — clickhouse/clickhouse-server, ports 8123/9000, volume at /var/lib/clickhouse. Langfuse v4 requires >= 25.12, 26.4 recommended.
• Postgres — managed PostgreSQL 18 (15+ required): accounts, projects, keys, prompts, datasets.
• Redis — managed Redis 8.2.1 (7+ required): BullMQ queues and the prompt cache.
• Object storage bucket — Railway managed S3 storage: one bucket, three prefixes (events/, media/, exports/). Its single endpoint serves server writes and presigned browser uploads, replacing MinIO.

Langfuse Environment Variables Reference

SALT and ENCRYPTION_KEY must be stable and identical on both; NEXTAUTH_SECRET is web-only.

Deployment Dependencies

langfuse/langfuse (MIT core, ee/ commercial); docs at langfuse.com/self-hosting. Pin both images to the same registry-verified tag — v4.2.0 was tagged but never published.

Hardware Requirements for Self-Hosting Langfuse

Idle draw is ~1.0–1.6 GB per container. NODE_OPTIONS=--max-old-space-size=3072 caps the Node heap; without it Node sizes it from the host's RAM and risks OOM.

Self-Hosting Langfuse with Docker

Upstream ships a Compose file covering both apps and all four stores:

git clone https://github.com/langfuse/langfuse.git
cd langfuse
docker compose up -d

Or run the app container alone against existing stores:

docker run -d --name langfuse-web -p 3000:3000 \
  -e DATABASE_URL="postgresql://user:pass@postgres:5432/langfuse" \
  -e CLICKHOUSE_URL="http://clickhouse:8123" \
  -e CLICKHOUSE_MIGRATION_URL="clickhouse://clickhouse:9000" \
  -e CLICKHOUSE_USER="langfuse" -e CLICKHOUSE_PASSWORD="" \
  -e REDIS_CONNECTION_STRING="redis://:pass@redis:6379" \
  -e NEXTAUTH_URL="https://langfuse.example.com" \
  -e NEXTAUTH_SECRET="" -e SALT="" \
  -e ENCRYPTION_KEY="<64 hex>" langfuse/langfuse:4.3.1

Run the worker image with the same variables and PORT=3030.

Langfuse vs LangSmith and Helicone

LangSmith suits all-LangChain stacks; Helicone wins on setup speed. Arize Phoenix, Braintrust and W&B Weave overlap on evals.

How Much Does Langfuse Cost to Self-Host?

Langfuse core is MIT-licensed and free to self-host with unlimited events and users; only the ee/ directory is commercially licensed, and those terms survived ClickHouse's acquisition of Langfuse in January 2026. Langfuse Cloud runs from a free Hobby tier (50k units/month) through Core at $29/month and Pro at $199/month to Enterprise at $2,499/month. Self-hosting on Railway costs infrastructure only.

Troubleshooting Self-Hosted Langfuse

ClickHouse logs Listen [0.0.0.0]:8123 failed: Address already in use → harmless. The image binds :: dual-stack first, so later IPv4 binds return EADDRINUSE, on 9000, 9004, 9005 and 9009 too. Confirm with Ready for connections.

ClickHouse logs are empty → the image logs to files, not stdout. Use the start command /entrypoint.sh -- --logger.console=1 --logger.level=information; the default trace level floods Railway's log cap.

langfuse-web logs dial tcp ...:9000: connection refused → web runs the ClickHouse migrations at boot and Railway has no start ordering, so a joint redeploy races ClickHouse. It retries and self-heals.

Browser media uploads fail → point LANGFUSE_S3_MEDIA_UPLOAD_ENDPOINT at the bucket's public endpoint and keep FORCE_PATH_STYLE=true.

FAQ

What is Langfuse?

An open-source, MIT-licensed platform for tracing, evaluating and managing prompts in AI apps. Every LLM call becomes a nested trace carrying tokens, cost and latency.

Why does self-hosted Langfuse need ClickHouse, Postgres, Redis and object storage?

Postgres stores accounts, projects, keys and prompts; ClickHouse is the OLAP store behind fast dashboards over millions of spans; Redis carries the ingestion queue and cache; the bucket holds raw events, media and exports.

What are the default login credentials for self-hosted Langfuse?

None. The first account registered becomes the owner, so create yours immediately, then set AUTH_DISABLE_SIGNUP=true.

How do I send OpenTelemetry traces to self-hosted Langfuse on Railway?

Point your OTLP HTTP exporter at /api/public/otel/v1/traces on your Railway domain, with HTTP Basic auth: public key as username, secret key as password.