Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/langfuse-pack/skills/langfuse-sdk-patterns/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill langfuse-sdk-patterns. Copyright stays with the author (MIT).

Langfuse SDK Patterns

Overview

Production-quality patterns for the Langfuse SDK: singleton clients, the observe wrapper, startActiveObservation for nested traces, session tracking, graceful shutdown, and error-safe tracing.

Prerequisites

• Completed langfuse-install-auth setup
• Understanding of async/await patterns
• For v4+: @langfuse/tracing, @langfuse/otel, @opentelemetry/sdk-node

Instructions

Pattern 1: Singleton Client with Graceful Shutdown

// src/lib/langfuse.ts -- single file, import everywhere
import { LangfuseClient } from "@langfuse/client";
import { LangfuseSpanProcessor } from "@langfuse/otel";
import { NodeSDK } from "@opentelemetry/sdk-node";

// Singleton client for prompts, datasets, scores
let client: LangfuseClient | null = null;
export function getLangfuseClient(): LangfuseClient {
  if (!client) {
    client = new LangfuseClient();
  }
  return client;
}

// One-time OTel setup (call at app entry point)
let sdk: NodeSDK | null = null;
export function initTracing(): NodeSDK {
  if (!sdk) {
    sdk = new NodeSDK({
      spanProcessors: [new LangfuseSpanProcessor()],
    });
    sdk.start();

    // Graceful shutdown on process exit
    const shutdown = async () => {
      await sdk?.shutdown();
      process.exit(0);
    };
    process.on("SIGTERM", shutdown);
    process.on("SIGINT", shutdown);
  }
  return sdk;
}

Legacy v3 singleton:

import { Langfuse } from "langfuse";

let instance: Langfuse | null = null;

export function getLangfuse(): Langfuse {
  if (!instance) {
    instance = new Langfuse({
      flushAt: 15,
      flushInterval: 10000,
    });
    process.on("beforeExit", () => instance?.shutdownAsync());
  }
  return instance;
}

Pattern 2: observe Wrapper for Existing Functions

The observe wrapper is the most ergonomic way to add tracing. It wraps any function and auto-creates a span.

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

// Wrap existing functions -- no internal changes needed
const fetchUserProfile = observe(async (userId: string) => {
  updateActiveObservation({ input: { userId } });
  const profile = await db.users.findById(userId);
  updateActiveObservation({ output: { found: !!profile } });
  return profile;
});

// Mark LLM calls as generations
const summarize = observe(
  { name: "summarize-text", asType: "generation" },
  async (text: string) => {
    updateActiveObservation({ model: "gpt-4o-mini", input: text });
    const result = await openai.chat.completions.create({
      model: "gpt-4o-mini",
      messages: [{ role: "user", content: `Summarize: ${text}` }],
    });
    const output = result.choices[0].message.content;
    updateActiveObservation({
      output,
      usage: {
        promptTokens: result.usage?.prompt_tokens,
        completionTokens: result.usage?.completion_tokens,
      },
    });
    return output;
  }
);

// When called inside another observed function, spans auto-nest
const pipeline = observe(async (userId: string) => {
  const profile = await fetchUserProfile(userId);
  const summary = await summarize(profile.bio);
  return { profile, summary };
});

Pattern 3: startActiveObservation for Inline Control

Use when you need fine-grained control over observation lifecycle within a function:

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

async function processOrder(orderId: string) {
  return await startActiveObservation("process-order", async () => {
    updateActiveObservation({ input: { orderId } });

    // Nested spans are automatic
    const validated = await startActiveObservation("validate", async () => {
      const result = await validateOrder(orderId);
      updateActiveObservation({ output: { valid: result.valid } });
      return result;
    });

    if (!validated.valid) {
      updateActiveObservation({ output: { error: "validation failed" } });
      return { success: false };
    }

    // Generation span for LLM call
    const description = await startActiveObservation(
      { name: "generate-confirmation", asType: "generation" },
      async () => {
        updateActiveObservation({ model: "gpt-4o-mini" });
        const result = await generateConfirmation(orderId);
        updateActiveObservation({ output: result });
        return result;
      }
    );

    updateActiveObservation({ output: { success: true } });
    return { success: true, description };
  });
}

Pattern 4: Session and User Tracking

Link traces across conversation turns for user-level analytics:

// v4+: Set session/user via observation metadata
await startActiveObservation("chat-turn", async () => {
  updateActiveObservation({
    metadata: {
      sessionId: "session-abc-123",
      userId: "user-456",
    },
  });
  // All nested observations inherit this context
  await handleUserMessage(message);
});

// v3: Set directly on trace
const trace = langfuse.trace({
  name: "chat-turn",
  sessionId: "session-abc-123", // Groups traces into a session
  userId: "user-456",           // Links to user analytics
  input: { message },
});

Pattern 5: Error-Safe Tracing

Never let tracing failures break your application:

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

const safeObserve = <T extends (...args: any[]) => Promise<any>>(
  name: string,
  fn: T
): T => {
  return (async (...args: Parameters<T>) => {
    try {
      return await observe({ name }, async () => {
        updateActiveObservation({ input: args });
        const result = await fn(...args);
        updateActiveObservation({ output: result });
        return result;
      })();
    } catch (tracingError) {
      // If tracing fails, still run the function
      console.warn(`Tracing error in ${name}:`, tracingError);
      return fn(...args);
    }
  }) as T;
};

// Usage -- function works even if Langfuse is down
const processRequest = safeObserve("process-request", async (input: string) => {
  return await callLLM(input);
});

Pattern 6: Legacy v3 -- Always End Spans

// Always use try/finally to ensure .end() is called
const span = trace.span({ name: "risky-operation", input: data });
try {
  const result = await riskyOperation(data);
  span.end({ output: result });
  return result;
} catch (error) {
  span.end({ level: "ERROR", statusMessage: String(error) });
  throw error;
}

Anti-Patterns to Avoid

Output

Use this skill to produce a small integration plan and implementation diff: one process-lifetime client, trace and observation names that describe the user operation, safe input/output capture, and a shutdown path that drains pending events. Record the SDK generation assumed by the code so a future upgrade does not mix v3 span APIs with the v4+ OpenTelemetry APIs.

Error Handling

Treat observability as non-critical infrastructure. A tracing export failure must be logged with enough context to diagnose it, but it must not replace or mask the application result. Avoid retrying a failed user operation merely to emit telemetry; instead rely on the SDK queue and alert on sustained export failures. Confirm legacy spans are ended in a finally block so an exception does not leave a permanently in-progress trace.

Examples

For a request handler, create an active observation named checkout-request, attach a bounded summary of the request, call the business function once, then attach its result summary. On process shutdown, await the tracing SDK shutdown before closing the HTTP server. This gives one trace per request without adding an extra network call to the request's success path.

Resources

• TypeScript SDK Instrumentation
• Advanced SDK Configuration
• Python Decorators
• Event Queuing/Batching

Next Steps

For OpenAI/LangChain tracing examples, see langfuse-core-workflow-a.