Langfuse Reference Architecture

Overview

Production-grade architecture patterns for Langfuse LLM observability: singleton SDK, context propagation with AsyncLocalStorage, cross-service trace correlation, multi-environment configurations, and scale strategies.

Prerequisites

• Understanding of distributed systems and async patterns
• Node.js 18+ with OpenTelemetry SDK
• For v4+: @langfuse/tracing, @langfuse/otel, @opentelemetry/sdk-node

Architecture Tiers

Instructions

Pattern 1: Singleton SDK with Context Propagation

// src/lib/tracing.ts -- Single module for all tracing
import { LangfuseClient } from "@langfuse/client";
import { LangfuseSpanProcessor } from "@langfuse/otel";
import { NodeSDK } from "@opentelemetry/sdk-node";
import { AsyncLocalStorage } from "async_hooks";

// Singleton OTel SDK
let sdk: NodeSDK | null = null;

export function initTracing() {
  if (sdk) return sdk;

  sdk = new NodeSDK({
    spanProcessors: [
      new LangfuseSpanProcessor({
        exportIntervalMillis: 5000,
        maxExportBatchSize: 50,
      }),
    ],
  });
  sdk.start();

  // Graceful shutdown
  for (const signal of ["SIGTERM", "SIGINT"]) {
    process.on(signal, async () => {
      console.log(`Received ${signal}, flushing traces...`);
      await sdk?.shutdown();
      process.exit(0);
    });
  }

  return sdk;
}

// Singleton client for non-tracing operations
let client: LangfuseClient | null = null;

export function getLangfuseClient(): LangfuseClient {
  if (!client) client = new LangfuseClient();
  return client;
}

// Request context for user/session tracking
interface RequestContext {
  userId?: string;
  sessionId?: string;
  requestId: string;
}

const requestStore = new AsyncLocalStorage<RequestContext>();

export function getRequestContext(): RequestContext | undefined {
  return requestStore.getStore();
}

export function runWithContext<T>(ctx: RequestContext, fn: () => T): T {
  return requestStore.run(ctx, fn);
}

Pattern 2: Express Middleware for Automatic Tracing

// src/middleware/tracing.ts
import { startActiveObservation, updateActiveObservation } from "@langfuse/tracing";
import { runWithContext, getRequestContext } from "../lib/tracing";
import { randomUUID } from "crypto";
import type { Request, Response, NextFunction } from "express";

export function langfuseMiddleware() {
  return (req: Request, res: Response, next: NextFunction) => {
    const ctx = {
      requestId: req.headers["x-request-id"]?.toString() || randomUUID(),
      userId: req.headers["x-user-id"]?.toString(),
      sessionId: req.headers["x-session-id"]?.toString(),
    };

    runWithContext(ctx, () => {
      startActiveObservation(`${req.method} ${req.path}`, async () => {
        updateActiveObservation({
          input: {
            method: req.method,
            path: req.path,
            query: req.query,
          },
          metadata: {
            userId: ctx.userId,
            sessionId: ctx.sessionId,
            requestId: ctx.requestId,
          },
        });

        // Capture response
        const originalEnd = res.end.bind(res);
        res.end = function (...args: any[]) {
          updateActiveObservation({
            output: { statusCode: res.statusCode },
          });
          return originalEnd(...args);
        } as any;

        next();
      }).catch(next);
    });
  };
}

// Usage
import express from "express";
import { initTracing } from "./lib/tracing";
import { langfuseMiddleware } from "./middleware/tracing";

initTracing();
const app = express();
app.use(langfuseMiddleware());

Pattern 3: Cross-Service Trace Correlation

For microservices, propagate trace context via HTTP headers:

// Service A: Inject trace context into outbound requests
import { context, propagation } from "@opentelemetry/api";

async function callServiceB(data: any) {
  const headers: Record<string, string> = {};

  // OTel propagation injects traceparent header automatically
  propagation.inject(context.active(), headers);

  const response = await fetch("https://service-b.internal/api/process", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      ...headers, // Includes traceparent, tracestate
    },
    body: JSON.stringify(data),
  });

  return response.json();
}

// Service B: Extract and continue trace context
import { context, propagation } from "@opentelemetry/api";
import { startActiveObservation, updateActiveObservation } from "@langfuse/tracing";

app.post("/api/process", async (req, res) => {
  // OTel automatically extracts context from incoming headers
  // when using standard HTTP instrumentation.
  // Any startActiveObservation call will be a child of the extracted trace.

  await startActiveObservation("service-b-process", async () => {
    updateActiveObservation({ input: req.body });
    const result = await processData(req.body);
    updateActiveObservation({ output: result });
    res.json(result);
  });
});

Pattern 4: Multi-Environment Configuration

// src/config/langfuse.ts
type Environment = "development" | "staging" | "production";

const configs: Record<Environment, {
  exportIntervalMillis: number;
  maxExportBatchSize: number;
  sampleRate: number;
}> = {
  development: {
    exportIntervalMillis: 1000,   // Immediate visibility
    maxExportBatchSize: 1,
    sampleRate: 1.0,              // Trace everything
  },
  staging: {
    exportIntervalMillis: 5000,
    maxExportBatchSize: 25,
    sampleRate: 0.5,              // 50% sampling
  },
  production: {
    exportIntervalMillis: 10000,
    maxExportBatchSize: 100,
    sampleRate: 0.1,              // 10% sampling
  },
};

// Pass the selected environment from the application's configuration boundary.
// Keeping configuration resolution outside tracing makes this module deterministic
// and straightforward to test.
export function getTracingConfig(env: Environment = "development") {
  return configs[env] || configs.development;
}

Pattern 5: Graceful Degradation

When Langfuse is unavailable, the app must keep running:

// The v4+ SDK with OTel handles this gracefully:
// - Failed exports are logged but don't throw
// - Events are buffered in the queue
// - Queue drops oldest events when maxQueueSize is exceeded
//
// For additional safety at the application level:

import { observe, updateActiveObservation } from "@langfuse/tracing";

let tracingHealthy = true;
let consecutiveFailures = 0;
const MAX_FAILURES = 10;

export function safeTrace<T extends (...args: any[]) => Promise<any>>(
  name: string,
  fn: T
): T {
  return (async (...args: Parameters<T>) => {
    if (!tracingHealthy) {
      return fn(...args); // Circuit breaker open
    }

    try {
      const result = await observe({ name }, async () => {
        updateActiveObservation({ input: args });
        const r = await fn(...args);
        updateActiveObservation({ output: r });
        return r;
      })();
      consecutiveFailures = 0;
      return result;
    } catch (error) {
      consecutiveFailures++;
      if (consecutiveFailures >= MAX_FAILURES) {
        tracingHealthy = false;
        console.error("Langfuse tracing disabled (circuit breaker open)");
        // Re-enable after 5 minutes
        setTimeout(() => { tracingHealthy = true; consecutiveFailures = 0; }, 300000);
      }
      return fn(...args);
    }
  }) as T;
}

Architecture Decision Matrix

Error Handling

Output

Produce an architecture decision record identifying the selected deployment tier, tracing boundary, context-propagation method, failure mode, retention owner, and rollback path. Include the tested revision and state whether the evidence comes from a local, staging, or production-like environment.

Examples

A growth-stage service can keep the SDK's internal queue and add AsyncLocalStorage propagation, then test that a request and downstream worker share the trace context. An enterprise deployment can put a queue between the application and self-hosted Langfuse, trip the circuit breaker during an export failure drill, and prove application traffic continues while telemetry recovers.

Resources

• TypeScript SDK Overview
• Advanced Configuration
• Self-Hosting Guide
• OpenTelemetry Context Propagation