langchain-callback-parquet-logger 0.3.0

A Parquet-based callback handler for logging LangChain LLM interactions

LangChain Callback Parquet Logger

A high-performance callback handler for logging LangChain LLM interactions to Parquet files with automatic partitioning, buffering, and batch processing support.

Features

• 📊 Parquet Format: Efficient columnar storage for analytics
• 🚀 Buffered Writing: Configurable buffer size for optimal performance
• 📅 Partitioning: Optional daily partitioning for better organization
• 🏷️ Custom Tracking: Add custom IDs and metadata to your logs
• 🔄 Batch Processing: Simple helper for DataFrame batch operations
• 🔒 Thread-Safe: Safe for concurrent LLM calls

Installation

pip install langchain-callback-parquet-logger

Quick Start

from langchain_callback_parquet_logger import ParquetLogger
from langchain_openai import ChatOpenAI

# Simple usage
llm = ChatOpenAI(model="gpt-4o-mini")
llm.callbacks = [ParquetLogger("./logs")]

response = llm.invoke("What is 2+2?")

Core Features

1. Basic Logging

# With context manager (recommended for notebooks)
with ParquetLogger('./logs') as logger:
    llm = ChatOpenAI(model="gpt-4o-mini", callbacks=[logger])
    response = llm.invoke("Hello!")
# Logs automatically flushed on exit

2. Custom IDs and Metadata

from langchain_callback_parquet_logger import ParquetLogger, with_tags

# Logger-level metadata (included in all logs)
logger = ParquetLogger(
    log_dir="./logs",
    logger_metadata={
        "environment": "production",
        "service": "api-gateway",
        "version": "2.1.0"
    }
)

# Request-level tracking with custom IDs
llm = ChatOpenAI(model="gpt-4o-mini", callbacks=[logger])
response = llm.invoke(
    "What is quantum computing?",
    config=with_tags(
        custom_id="user-123-req-456",
        tags=["production", "high-priority"]
    )
)

3. Batch Processing (v0.3.0+)

Process DataFrames through LLMs with minimal code:

import pandas as pd
from langchain_callback_parquet_logger import batch_run, with_tags, ParquetLogger

# Prepare your data
df = pd.DataFrame({
    'id': ['001', '002', '003'],
    'question': ['What is AI?', 'Explain quantum computing', 'What is blockchain?']
})

# Add required columns
df['prompt'] = df['question']
df['config'] = df['id'].apply(lambda x: with_tags(custom_id=x))

# Configure LLM with advanced features
with ParquetLogger('./logs') as logger:
    llm = ChatOpenAI(
        model="gpt-4o-mini",
        service_tier="flex",  # Optional: optimize costs
        model_kwargs={"background": True},  # Optional: background processing
        callbacks=[logger]
    )
    
    # Run batch processing
    results = await batch_run(df, llm, max_concurrency=10, show_progress=True)
    df['answer'] = results

See examples/batch_processing.py for advanced usage with structured outputs, web search tools, and more.

Memory-Efficient Mode for Huge DataFrames

For massive DataFrames, use return_results=False to avoid keeping results in memory:

# Process huge DataFrame without memory overhead
with ParquetLogger('./logs') as logger:
    llm = ChatOpenAI(model="gpt-4o-mini", callbacks=[logger])
    
    # Results saved to parquet only, not kept in memory
    await batch_run(huge_df, llm, return_results=False)
    
# Read results later from parquet files
df_logs = pd.read_parquet('./logs')
results = df_logs[df_logs['event_type'] == 'llm_end']

Configuration Options

ParquetLogger Parameters

Parameter	Type	Default	Description
log_dir	str	"./llm_logs"	Directory for log files
buffer_size	int	100	Entries before auto-flush
provider	str	"openai"	LLM provider name
logger_metadata	dict	{}	Metadata for all logs
partition_on	str/None	"date"	"date" or None for no partitioning

batch_run Parameters

Parameter	Type	Default	Description
df	DataFrame	required	DataFrame with data
llm	LangChain LLM	required	Configured LLM instance
prompt_col	str	"prompt"	Column with prompts
config_col	str	"config"	Column with config dicts
tools_col	str/None	"tools"	Column with tools lists
max_concurrency	int	10	Max parallel requests
show_progress	bool	True	Show progress bar
return_results	bool	True	If False, don't keep results in memory

Reading Logs

With Pandas

import pandas as pd
import json

df = pd.read_parquet("./logs")
df['data'] = df['payload'].apply(json.loads)

# Analyze by custom ID
custom_requests = df[df['logger_custom_id'] != '']
print(f"Found {len(custom_requests)} tagged requests")

With DuckDB

import duckdb

conn = duckdb.connect()
df = conn.execute("""
    SELECT 
        logger_custom_id,
        event_type,
        timestamp,
        json_extract_string(payload, '$.usage.total_tokens') as tokens
    FROM read_parquet('./logs/**/*.parquet')
    WHERE logger_custom_id != ''
    ORDER BY timestamp DESC
""").df()

Log Schema

Column	Type	Description
timestamp	timestamp	Event time (UTC)
run_id	string	Unique run ID
logger_custom_id	string	Your custom ID
event_type	string	llm_start/end/error
provider	string	LLM provider
logger_metadata	string	JSON metadata
payload	string	JSON event data

Important Notes

Notebook Usage

In Jupyter/Colab, use one of these approaches for immediate writes:

• Context manager (recommended): with ParquetLogger() as logger:
• Small buffer: ParquetLogger(buffer_size=1)
• Manual flush: logger.flush()

File Organization

logs/
├── date=2024-01-15/          # With partitioning (default)
│   └── logs_143022_123456.parquet
└── date=2024-01-16/
    └── logs_090122_345678.parquet

Examples

• basic_usage.py - Simple logging example
• batch_processing.py - Advanced batch processing with all features
• simple_batch_example.py - Before/after batch processing comparison
• memory_efficient_batch.py - Memory-efficient processing for huge DataFrames
• partitioning_example.py - Partitioning strategies

License

MIT License - see LICENSE file

Contributing

Contributions welcome! Please submit a Pull Request.

Support

For issues and questions, use GitHub issues.