Self-hosting

Run langprobe in your VPC — one compose file for local, a Helm chart and operator for Kubernetes.

langprobe is self-hosted by design: it runs in your VPC, on your data, under Apache-2.0. The whole stack is one docker compose file locally, and a Helm chart plus an operator for Kubernetes.

Architecture

Four stores, each purpose-built (see Core concepts for the data model):

Postgres — control plane: orgs, users, projects, API keys, audit log.
ClickHouse — data plane: runs, spans, eval scores, replay captures.
Redis — the ingest queue (plus rate-limit buckets and cache).
Object storage (S3/MinIO) — large attachments and inputs.

Services in front of them: an ingest-api (OTLP + LangSmith-shim intake), an ingest-worker (drains Redis → ClickHouse), the api (auth, CRUD, replay), and the web UI.

Local: docker compose

cp infra/.env.example infra/.env
echo "SESSION_SECRET=$(openssl rand -hex 32)" >> infra/.env
docker compose -f infra/docker-compose.yml up --build

Brings up Postgres, ClickHouse, Redis, and the services on ports 7080 (ingest), 7081 (api), and 7090 (web). See the Quickstart for the first-run walkthrough.

Kubernetes: Helm & operator

For production, deploy/ ships a Helm chart, raw k8s manifests, and an operator so langprobe drops into your cluster the same way the incumbents' charts do — but in your VPC.

Configuration

Configuration is environment variables, with no silent defaults for secrets — missing required values fail loudly at startup rather than defaulting:

SESSION_SECRET (≥ 32 chars) — required; generate with openssl rand -hex 32.
LANGPROBE_PG_DSN — Postgres connection string.
LANGPROBE_CLICKHOUSE_URL — ClickHouse endpoint.
LANGPROBE_REDIS_URL — Redis URL for the ingest queue.

Where things live

Useful when something looks stuck:

Redis stream langprobe:ingest:v1 (consumer group ingest); a dead-letter queue langprobe:ingest:v1:dlq after repeated redeliveries.
ClickHouse tables — run, span, eval_score, eval_aggregate, replay_capture, replay_run.
Postgres tables — app_user, org, workspace, project, api_key, audit_log. Every state-changing call leaves an audit_log row.

Because ingest is a queue append, POST /v1/traces and /v1/runs return 202 Accepted immediately and the worker writes to ClickHouse asynchronously — so a backlog shows up as Redis stream depth, not dropped traces.