LangChain Incident Runbook

Overview

Standard operating procedures for responding to LangChain production incidents with diagnosis, mitigation, and recovery steps.

Prerequisites

• Access to production infrastructure
• Monitoring dashboards configured
• LangSmith or equivalent tracing
• On-call rotation established

Incident Classification

Severity Levels

Runbook: LLM Provider Outage

Detection

# Check if LLM provider is responding
curl -s https://status.openai.com/api/v2/status.json | jq '.status.indicator'
curl -s https://status.anthropic.com/api/v2/status.json | jq '.status.indicator'

# Check your error rate
# Prometheus query:
# sum(rate(langchain_llm_requests_total{status="error"}[5m])) / sum(rate(langchain_llm_requests_total[5m]))

Diagnosis

# Quick diagnostic script
import asyncio
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

async def diagnose_providers():
    """Check all configured providers."""
    results = {}

    # Test OpenAI
    try:
        llm = ChatOpenAI(model="gpt-4o-mini", request_timeout=10)
        await llm.ainvoke("test")
        results["openai"] = "OK"
    except Exception as e:
        results["openai"] = f"FAIL: {e}"

    # Test Anthropic
    try:
        llm = ChatAnthropic(model="claude-3-5-sonnet-20241022", timeout=10)
        await llm.ainvoke("test")
        results["anthropic"] = "OK"
    except Exception as e:
        results["anthropic"] = f"FAIL: {e}"

    return results

# Run
print(asyncio.run(diagnose_providers()))

Mitigation: Enable Fallback

# Emergency fallback configuration
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

# Original
llm = ChatOpenAI(model="gpt-4o-mini")

# With fallback
primary = ChatOpenAI(model="gpt-4o-mini", max_retries=1, request_timeout=5)
fallback = ChatAnthropic(model="claude-3-haiku-20240307")

llm = primary.with_fallbacks([fallback])

Recovery

1. Monitor provider status page
2. Gradually remove fallback when primary recovers
3. Document incident in post-mortem

Runbook: High Error Rate

Detection

# Check recent errors in logs
grep -i "error" /var/log/langchain/app.log | tail -50

# Check LangSmith for failed traces
# Navigate to: https://smith.langchain.com/o/YOUR_ORG/projects/YOUR_PROJECT/runs?filter=error%3Atrue

Diagnosis

# Analyze error distribution
from collections import Counter
import json

def analyze_errors(log_file: str) -> dict:
    """Analyze error patterns from logs."""
    errors = []

    with open(log_file) as f:
        for line in f:
            if "error" in line.lower():
                try:
                    log = json.loads(line)
                    errors.append(log.get("error_type", "unknown"))
                except:
                    pass

    return dict(Counter(errors).most_common(10))

# Common error types and causes
ERROR_CAUSES = {
    "RateLimitError": "Exceeded API quota - reduce load or increase limits",
    "AuthenticationError": "Invalid API key - check secrets",
    "Timeout": "Network issues or overloaded provider",
    "OutputParserException": "LLM output format changed - check prompts",
    "ValidationError": "Schema mismatch - update Pydantic models",
}

Mitigation

# 1. Reduce load
# Scale down instances or enable circuit breaker

# 2. Emergency rate limiting
from functools import wraps
import time

def emergency_rate_limit(calls_per_minute: int = 10):
    """Emergency rate limiter decorator."""
    interval = 60.0 / calls_per_minute
    last_call = [0]

    def decorator(func):
        @wraps(func)
        async def wrapper(*args, **kwargs):
            elapsed = time.time() - last_call[0]
            if elapsed < interval:
                await asyncio.sleep(interval - elapsed)
            last_call[0] = time.time()
            return await func(*args, **kwargs)
        return wrapper
    return decorator

# 3. Enable caching for repeated queries
from langchain_core.globals import set_llm_cache
from langchain_community.cache import InMemoryCache
set_llm_cache(InMemoryCache())

Runbook: Memory/Performance Issues

Detection

# Check memory usage
ps aux | grep python | head -5

# Check for memory leaks
# Prometheus: process_resident_memory_bytes

Diagnosis

# Memory profiling
import tracemalloc

tracemalloc.start()

# Run your chain
chain.invoke({"input": "test"})

snapshot = tracemalloc.take_snapshot()
top_stats = snapshot.statistics('lineno')

print("Top 10 memory allocations:")
for stat in top_stats[:10]:
    print(stat)

Mitigation

# 1. Clear caches
from langchain_core.globals import set_llm_cache
set_llm_cache(None)

# 2. Reduce batch sizes
# Change from: chain.batch(inputs, config={"max_concurrency": 50})
# To: chain.batch(inputs, config={"max_concurrency": 10})

# 3. Restart pods gracefully
# kubectl rollout restart deployment/langchain-api

Runbook: Cost Spike

Detection

# Check token usage
# Prometheus: sum(increase(langchain_llm_tokens_total[1h]))

# OpenAI usage dashboard
# https://platform.openai.com/usage

Diagnosis

# Identify high-cost operations
def analyze_costs(traces: list) -> dict:
    """Analyze cost from trace data."""
    by_chain = {}

    for trace in traces:
        chain_name = trace.get("name", "unknown")
        tokens = trace.get("total_tokens", 0)

        if chain_name not in by_chain:
            by_chain[chain_name] = {"count": 0, "tokens": 0}

        by_chain[chain_name]["count"] += 1
        by_chain[chain_name]["tokens"] += tokens

    return sorted(by_chain.items(), key=lambda x: x[1]["tokens"], reverse=True)

Mitigation

# 1. Emergency budget limit
class BudgetExceeded(Exception):
    pass

daily_spend = 0
DAILY_LIMIT = 100.0  # $100

def check_budget(cost: float):
    global daily_spend
    daily_spend += cost
    if daily_spend > DAILY_LIMIT:
        raise BudgetExceeded(f"Daily limit ${DAILY_LIMIT} exceeded")

# 2. Switch to cheaper model
# gpt-4o -> gpt-4o-mini (30x cheaper)
# claude-3-5-sonnet -> claude-3-haiku (12x cheaper)

# 3. Enable aggressive caching

Incident Response Checklist

During Incident

• Acknowledge incident in Slack/PagerDuty
• Identify severity level
• Start incident channel/call
• Begin diagnosis
• Implement mitigation
• Communicate status to stakeholders
• Document timeline

Post-Incident

• Verify full recovery
• Update status page
• Schedule post-mortem
• Write incident report
• Create follow-up tickets
• Update runbooks

Resources

• OpenAI Status
• Anthropic Status
• LangSmith
• PagerDuty Best Practices

Next Steps

Use langchain-debug-bundle for detailed evidence collection.