Langfuse Performance Tuning

Langfuse Performance Tuning

Overview

Optimize Langfuse tracing for minimal overhead and maximum throughput.

Prerequisites

• Existing Langfuse integration
• Performance baseline measurements
• Understanding of async patterns

Performance Targets

Instructions

Step 1: Measure Baseline Performance

// scripts/benchmark-langfuse.ts
import { Langfuse } from "langfuse";
import { performance } from "perf_hooks";

async function benchmark() {
  const langfuse = new Langfuse();
  const iterations = 1000;  # 1000: 1 second in ms

  // Measure trace creation
  const traceTimings: number[] = [];
  for (let i = 0; i < iterations; i++) {
    const start = performance.now();
    const trace = langfuse.trace({ name: `benchmark-${i}` });
    traceTimings.push(performance.now() - start);
  }

  // Measure generation creation
  const genTimings: number[] = [];
  const trace = langfuse.trace({ name: "gen-benchmark" });
  for (let i = 0; i < iterations; i++) {
    const start = performance.now();
    const gen = trace.generation({
      name: `gen-${i}`,
      model: "gpt-4",
      input: [{ role: "user", content: "test" }],
    });
    gen.end({ output: "response" });
    genTimings.push(performance.now() - start);
  }

  // Measure flush
  const flushStart = performance.now();
  await langfuse.flushAsync();
  const flushTime = performance.now() - flushStart;

  // Calculate stats
  const stats = (arr: number[]) => ({
    mean: arr.reduce((a, b) => a + b) / arr.length,
    p50: arr.sort((a, b) => a - b)[Math.floor(arr.length * 0.5)],
    p95: arr.sort((a, b) => a - b)[Math.floor(arr.length * 0.95)],
    p99: arr.sort((a, b) => a - b)[Math.floor(arr.length * 0.99)],
  });

  console.log("=== Langfuse Performance Benchmark ===");
  console.log(`\nTrace Creation (${iterations} iterations):`);
  console.log(JSON.stringify(stats(traceTimings), null, 2));
  console.log(`\nGeneration Creation (${iterations} iterations):`);
  console.log(JSON.stringify(stats(genTimings), null, 2));
  console.log(`\nFlush Time: ${flushTime.toFixed(2)}ms`);

  await langfuse.shutdownAsync();
}

benchmark();

Step 2: Optimize Batching Configuration

// High-throughput configuration
const langfuse = new Langfuse({
  publicKey: process.env.LANGFUSE_PUBLIC_KEY!,
  secretKey: process.env.LANGFUSE_SECRET_KEY!,

  // Batching optimization
  flushAt: 100,           // Larger batches = fewer requests
  flushInterval: 10000,   // Less frequent flushes  # 10000: 10 seconds in ms

  // Timeout tuning
  requestTimeout: 30000,  // Allow time for large batches  # 30000: 30 seconds in ms

  // Optional: Disable in development
  enabled: process.env.NODE_ENV === "production",
});

Step 3: Implement Non-Blocking Tracing

// Trace wrapper that never blocks the main operation
class NonBlockingLangfuse {
  private langfuse: Langfuse;
  private errorCount = 0;
  private maxErrors = 10;

  constructor(config: ConstructorParameters<typeof Langfuse>[0]) {
    this.langfuse = new Langfuse(config);
  }

  // Fire-and-forget trace
  trace(params: Parameters<typeof this.langfuse.trace>[0]) {
    if (this.errorCount >= this.maxErrors) {
      // Circuit breaker: stop tracing if too many errors
      return this.createNoOpTrace();
    }

    try {
      return this.langfuse.trace(params);
    } catch (error) {
      this.errorCount++;
      console.error("Langfuse trace error:", error);
      return this.createNoOpTrace();
    }
  }

  private createNoOpTrace() {
    return {
      id: "noop",
      span: () => this.createNoOpSpan(),
      generation: () => this.createNoOpGeneration(),
      update: () => {},
      getTraceUrl: () => "",
    };
  }

  private createNoOpSpan() {
    return {
      id: "noop",
      span: () => this.createNoOpSpan(),
      generation: () => this.createNoOpGeneration(),
      end: () => {},
    };
  }

  private createNoOpGeneration() {
    return {
      id: "noop",
      end: () => {},
    };
  }

  // Background flush - don't await in hot path
  flush() {
    this.langfuse.flushAsync().catch((error) => {
      this.errorCount++;
      console.error("Langfuse flush error:", error);
    });
  }

  async shutdown() {
    return this.langfuse.shutdownAsync();
  }
}

Step 4: Optimize Data Payload Size

// Reduce trace payload size
function optimizeTraceInput(input: any): any {
  // Truncate large strings
  const MAX_STRING_LENGTH = 10000;  # 10000: 10 seconds in ms

  if (typeof input === "string") {
    return input.length > MAX_STRING_LENGTH
      ? input.slice(0, MAX_STRING_LENGTH) + "...[truncated]"
      : input;
  }

  if (Array.isArray(input)) {
    // Limit array size
    const MAX_ARRAY_LENGTH = 100;
    const truncated = input.slice(0, MAX_ARRAY_LENGTH);
    return truncated.map(optimizeTraceInput);
  }

  if (typeof input === "object" && input !== null) {
    // Remove large binary data
    const optimized: Record<string, any> = {};
    for (const [key, value] of Object.entries(input)) {
      if (value instanceof Buffer || value instanceof Uint8Array) {
        optimized[key] = `[Binary: ${value.length} bytes]`;
      } else {
        optimized[key] = optimizeTraceInput(value);
      }
    }
    return optimized;
  }

  return input;
}

// Use in traces
const trace = langfuse.trace({
  name: "optimized-trace",
  input: optimizeTraceInput(largeInput),
});

Step 5: Implement Sampling for Ultra-High Volume

interface SamplingStrategy {
  shouldSample(params: TraceParams): boolean;
}

// Deterministic sampling based on trace attributes
class DeterministicSampler implements SamplingStrategy {
  private rate: number;

  constructor(rate: number) {
    this.rate = rate;
  }

  shouldSample(params: TraceParams): boolean {
    // Always sample errors
    if (params.level === "ERROR" || params.tags?.includes("error")) {
      return true;
    }

    // Deterministic hash-based sampling
    const hash = this.hashString(params.name + (params.userId || ""));
    return (hash % 100) < (this.rate * 100);
  }

  private hashString(str: string): number {
    let hash = 0;
    for (let i = 0; i < str.length; i++) {
      const char = str.charCodeAt(i);
      hash = ((hash << 5) - hash) + char;
      hash = hash & hash;
    }
    return Math.abs(hash);
  }
}

// Adaptive sampling based on throughput
class AdaptiveSampler implements SamplingStrategy {
  private windowMs = 60000;  # 60000: 1 minute in ms
  private maxPerWindow = 1000;  # 1000: 1 second in ms
  private counts: number[] = [];

  shouldSample(params: TraceParams): boolean {
    const now = Date.now();
    const windowStart = now - this.windowMs;

    // Clean old counts
    this.counts = this.counts.filter((t) => t > windowStart);

    // Check if under limit
    if (this.counts.length < this.maxPerWindow) {
      this.counts.push(now);
      return true;
    }

    // Over limit - only sample important traces
    return params.level === "ERROR";
  }
}

// Apply sampling
const sampler = new DeterministicSampler(0.1); // 10% sampling

function sampledTrace(params: TraceParams) {
  if (!sampler.shouldSample(params)) {
    return createNoOpTrace();
  }

  return langfuse.trace({
    ...params,
    metadata: {
      ...params.metadata,
      sampled: true,
    },
  });
}

Step 6: Memory Management

// Prevent memory leaks with trace cleanup
class ManagedLangfuse {
  private langfuse: Langfuse;
  private activeTraces: Map<string, { createdAt: Date }> = new Map();
  private maxTraceAge = 300000; // 5 minutes  # 300000 = configured value

  constructor(config: ConstructorParameters<typeof Langfuse>[0]) {
    this.langfuse = new Langfuse(config);

    // Periodic cleanup
    setInterval(() => this.cleanupStaleTraces(), 60000);  # 60000: 1 minute in ms
  }

  trace(params: Parameters<typeof this.langfuse.trace>[0]) {
    const trace = this.langfuse.trace(params);
    this.activeTraces.set(trace.id, { createdAt: new Date() });
    return trace;
  }

  private cleanupStaleTraces() {
    const now = Date.now();
    let cleaned = 0;

    for (const [id, meta] of this.activeTraces) {
      if (now - meta.createdAt.getTime() > this.maxTraceAge) {
        this.activeTraces.delete(id);
        cleaned++;
      }
    }

    if (cleaned > 0) {
      console.log(`Cleaned up ${cleaned} stale trace references`);
    }
  }

  getStats() {
    return {
      activeTraces: this.activeTraces.size,
      heapUsed: process.memoryUsage().heapUsed / 1024 / 1024,  # 1024: 1 KB
    };
  }
}

Output

• Baseline performance measurements
• Optimized batching configuration
• Non-blocking trace wrapper
• Payload size optimization
• Sampling strategies for high volume
• Memory leak prevention

Performance Optimization Checklist

Error Handling

Resources

• Langfuse SDK Configuration
• Node.js Performance
• Sampling Strategies

Next Steps

For cost optimization, see langfuse-cost-tuning.

Examples

Basic usage: Apply langfuse performance tuning to a standard project setup with default configuration options.

Advanced scenario: Customize langfuse performance tuning for production environments with multiple constraints and team-specific requirements.