llm-gateway-sub-backup/workspace/utils/logging_utils.py

import csv
import json
import logging
import time
from datetime import datetime
from pathlib import Path

import redis
from config.setting import PGN_REDIS_DB, PGN_REDIS_HOST, PGN_REDIS_PORT
from fastapi import Depends, Request

from utils.checking_keys import get_api_key
from utils.request_utils import get_client_ip, get_swagger_port

logger = logging.getLogger(__name__)

redis_client = redis.Redis(
    host=PGN_REDIS_HOST, port=PGN_REDIS_PORT, db=PGN_REDIS_DB, decode_responses=True
)


def log_user_request(
    request_info: str,
    endpoint: str,
    input_filename: str,
    model: str,
    prompt_filename: str,
    context_length: int,
    api_key: str,
):
    client_ip = get_client_ip(request_info)
    swagger_port = get_swagger_port(request_info)

    # ✅ 1. CSV 파일 저장
    logs_dir = Path("./logs")
    logs_dir.mkdir(exist_ok=True)

    csv_path = logs_dir / f"{client_ip}.csv"
    new_file = not csv_path.exists()

    with open(csv_path, mode="a", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        if new_file:
            writer.writerow(
                [
                    "timestamp",
                    "swagger_port",
                    "endpoint",
                    "input_filename",
                    "prompt_filename",
                    "model",
                    "context_length",
                    "api_key",
                ]
            )
        writer.writerow(
            [
                time.strftime("%Y-%m-%d %H:%M:%S"),
                swagger_port,
                endpoint,
                input_filename,
                prompt_filename,
                model,
                context_length,
                api_key,
            ]
        )

    # ✅ 2. Loki용 JSON 로그 출력
    logger.info(
        json.dumps(
            {
                "ip": client_ip,
                "swagger_port": swagger_port,
                "endpoint": endpoint,
                "input_filename": input_filename,
                "prompt_filename": prompt_filename,
                "model": model,
                "context_length": context_length,
                "api_key": api_key,
                "event": "inference_log",
            }
        )
    )


def log_generation_info(custom_mode: bool, user_prompt: str = ""):
    logger.info(f"[GENERATE-PROMPT-USED] 사용자 정의 프롬프트 사용유무: {custom_mode}")

    if custom_mode:
        logger.info(f"[GENERATE-USER-PROMPT]\n{user_prompt}")
    else:
        logger.info("[GENERATE-DEFAULT-PROMPT] Default_prompt")


def log_pipeline_status(request_id: str, status_message: str, step_info: dict = None):
    log_entry = {
        "status": status_message,
        "timestamp": datetime.now().isoformat(),
        "step_info": step_info,
    }
    redis_client.rpush(f"pipeline_status:{request_id}", json.dumps(log_entry))


def ns_to_sec(ns: int) -> float:
    """나노초를 초로 변환"""
    return round(ns / 1e9, 3)  # 소수점 3자리


def log_ollama_stats(res: dict):
    """Ollama 응답 JSON 내 추론 통계를 한 줄 JSON 로그로 출력 (초 단위 변환 + token/s 포함)"""
    # 원본 값
    total_duration = res.get("total_duration")
    load_duration = res.get("load_duration")
    prompt_eval_count = res.get("prompt_eval_count")
    prompt_eval_duration = res.get("prompt_eval_duration")
    eval_count = res.get("eval_count")
    eval_duration = res.get("eval_duration")

    # 초 단위로 변환
    stats = {
        "model": res.get("model"),
        "total_duration_ns": total_duration,
        "total_duration_sec": ns_to_sec(total_duration),
        "load_duration_ns": load_duration,
        "load_duration_sec": ns_to_sec(load_duration),
        "prompt_eval_count": prompt_eval_count,
        "prompt_eval_duration_ns": prompt_eval_duration,
        "prompt_eval_duration_sec": ns_to_sec(prompt_eval_duration),
        "eval_count": eval_count,
        "eval_duration_ns": eval_duration,
        "eval_duration_sec": ns_to_sec(eval_duration),
    }

    # token/s 계산
    if eval_count and eval_duration:
        stats["generation_speed_tok_per_sec"] = round(
            eval_count / (eval_duration / 1e9), 2
        )

    logger.info("[OLLAMA-STATS] " + json.dumps(stats, ensure_ascii=False))


class EndpointLogger:
    def __init__(self, request: Request, api_key: str = Depends(get_api_key)):
        self.request = request
        self.api_key = api_key

    def log(
        self,
        model: str,
        input_filename: str = "N/A",
        prompt_filename: str = "N/A",
        context_length: int = 0,
    ):
        try:
            log_user_request(
                request_info=self.request,
                endpoint=self.request.url.path,
                input_filename=input_filename,
                model=model,
                prompt_filename=prompt_filename,
                context_length=context_length,
                api_key=self.api_key,
            )
        except Exception as e:
            logger.warning(
                f"Failed to log request for endpoint {self.request.url.path}: {e}"
            )

class HealthCheckFilter(logging.Filter):
    def filter(self, record: logging.LogRecord) -> bool:
        # The access log record for uvicorn has the data in `args`.
        # record.args = (client_addr, method, path, http_version, status_code)
        # e.g. ('127.0.0.1:37894', 'GET', '/health/API', '1.1', 200)
        if isinstance(record.args, tuple) and len(record.args) == 5:
            method = record.args[1]
            path = record.args[2]
            status_code = record.args[4]
            if method == 'GET' and isinstance(path, str) and path.startswith('/health') and status_code == 200:
                return False
        return True