Skip to main content

LangGuard on Databricks

LangGuard can be deployed natively inside your own Databricks workspace as a Databricks App. In this mode LangGuard runs entirely on your Databricks compute, stores all of its data in your Unity Catalog and Lakebase Postgres, and governs the agents and model endpoints running in your workspace — no data ever leaves your Databricks account.

This is different from the Databricks integration for the LangGuard SaaS, which forwards activity from a Databricks workspace to the hosted LangGuard service. If you need full data residency, the Databricks App is the right choice.

Single-tenant by design

A Databricks App deployment serves one tenant — your workspace. There is no multi-tenant routing; the app is provisioned, secured, and billed per workspace.

Why run LangGuard as a Databricks App​

  • Data residency — traces, policies, violations, and discovered entities live in your Unity Catalog and Lakebase, governed by your existing UC permissions.
  • Native AI Gateway enforcement — LangGuard can sit inline in front of your Mosaic AI Gateway / Model Serving endpoints and allow, notify, or block model calls in real time. See AI Gateway Enforcement.
  • Automatic capture — model and agent activity is captured from the workspace with no per-agent instrumentation, via native OTLP ingestion and AI Gateway inference tables. See Data Capture.
  • Unity Catalog governance — every LangGuard table is a real UC object you can query, secure, and audit. See Unity Catalog Data.

Architecture at a glance​

The app ships as a single, self-contained artifact that runs three coordinated processes inside the Databricks App container:

        ┌──────────────────────── Databricks App container ────────────────────────┐
        │                                                                           │
        │   LangGuard server (Deno-compiled binary)                                 │
        │     • Web UI + API            • Policy enforcement (/v1/chat proxy)        │
        │     • OTLP exporter           • GUI-aware pollers                          │
        │                                                                           │
        │   OPA sidecar                 OpenCITE sidecar                             │
        │     • Rego policy eval          • Agent / tool / model / MCP discovery     │
        │                                                                           │
        └───────────┬───────────────────────────────┬───────────────────────────────┘
                    │                                │
       Lakebase PostgreSQL                   Unity Catalog (Delta)
       (metadata, policies,                  (OTEL spans/logs, trace
        entities, config)                     views, annotations)

Unlike the LangGuard SaaS pipeline, the Databricks App has no otel-ingest service and no inbound webhook chain. Instead the app:

  1. Enforces model calls that route through its /v1/chat/completions proxy (the AI Gateway path), and
  2. Exports the resulting traces as OTLP directly to your workspace's native Databricks OTLP endpoint, which lands them in a Unity Catalog Delta table, and
  3. Polls those Delta tables (plus AI Gateway inference tables and system.ai_gateway.usage) to surface traces, run discovery, and evaluate policies.

How your model activity is captured​

LangGuard captures activity through three complementary paths so it can see everything in the workspace — even agents it doesn't directly proxy:

PathWhat it coversEnforcement
/v1/chat proxy (Guarded Endpoints)Calls routed through a LangGuard-wrapped AI Gateway endpointReal-time allow / notify / block
Native OTLP exportTraces LangGuard emits for proxied callsObserve
PollersAI Gateway inference tables + system.ai_gateway.usage for endpoints that don't proxyObserve

See Data Capture for the full flow.

Next steps​