""" 请求日志服务:usage_log 表的写入与统计 - record_usage_log: 请求完成时写入一条日志(不抛异常,不影响主流程) 写入先进入内存队列,由后台任务批量落库,避免在请求热路径上等待 SQLite 写锁 - extract_usage_tokens: 从 Gemini 响应/流式块中提取 usageMetadata token 计数 """ import asyncio import time import uuid from typing import Any, Dict, Optional import config from log import log from src.storage.database import get_database def new_request_id() -> str: """生成短请求 ID""" return uuid.uuid4().hex[:16] # ====================== 异步批量写入队列 ====================== # 内存队列:请求热路径只做 append(O(1)),后台任务批量写库 _pending_logs: list = [] _queue_lock = asyncio.Lock() # 后台批量刷新任务(应用启动后首次写入时惰性创建) _flush_task: Optional[asyncio.Task] = None # 批量落库的间隔(秒);缓冲区满时立即触发刷新 _FLUSH_INTERVAL = 2.0 # 缓冲区上限(超过立即批量落库,防止异常高峰下内存增长) _MAX_PENDING = 500 _FLUSH_LOOP_STOP = False async def _flush_pending_logs() -> None: """将待写日志批量落库(分批事务,单条失败不影响整批)""" global _pending_logs async with _queue_lock: batch = _pending_logs _pending_logs = [] if not batch: return database = await get_database() for entry in batch: try: await database.insert_usage_log(entry) except Exception as e: log.debug(f"请求日志写入失败(不影响主流程): {e}") async def _flush_loop() -> None: """后台定时批量落库循环""" while True: await asyncio.sleep(_FLUSH_INTERVAL) try: await _flush_pending_logs() except Exception as e: log.debug(f"请求日志批量落库异常: {e}") def _ensure_flush_loop() -> None: """确保后台刷新任务存在(惰性启动,随事件循环生命周期)""" global _flush_task if _flush_task is None or _flush_task.done(): _flush_task = asyncio.get_event_loop().create_task(_flush_loop()) async def flush_usage_logs() -> None: """立即落库所有待写日志(应用关闭时调用,避免丢日志)""" await _flush_pending_logs() def _enqueue_log(entry: Dict[str, Any]) -> None: """入队一条日志;缓冲区超限时同步触发一次批量落库""" global _pending_logs _pending_logs.append(entry) if len(_pending_logs) >= _MAX_PENDING: try: # 已在事件循环内:调度立即刷新(不阻塞请求) asyncio.get_event_loop().create_task(_flush_pending_logs()) except Exception: pass async def _write_log(entry: Dict[str, Any]) -> None: """日志写入入口:开关校验后进入批量队列""" try: if not await config.get_usage_log_enabled(): return _ensure_flush_loop() _enqueue_log({**entry, "created_at": entry.get("created_at") or time.time()}) except Exception as e: log.debug(f"请求日志入队失败(不影响主流程): {e}") def extract_usage_tokens(source: Any) -> Dict[str, int]: """ 从 Gemini 响应结构中提取 token 用量。 兼容: - 非流式: {"usageMetadata": {...}} 或 {"response": {"usageMetadata": {...}}} - 流式块: data 字符串内的 usageMetadata(通常只在最后一个块出现) Gemini 语义(与 PR#432 对齐): - promptTokenCount 含缓存 token(cachedContentTokenCount) - completion = candidatesTokenCount + thoughtsTokenCount 同时兼容 v1internal 新版 Interactions 格式(对齐 Antigravity-Manager CanonicalUsage::from_gemini): - total_input_tokens / total_output_tokens / total_thought_tokens / total_cached_tokens """ prompt_tokens = completion_tokens = total_tokens = cached_tokens = 0 if isinstance(source, (str, bytes)): # 流式 SSE 块:解析 JSON 提取 usageMetadata try: text = source.decode("utf-8") if isinstance(source, bytes) else source if text.startswith("data: "): text = text[len("data: "):].strip() if not text or text == "[DONE]": return {} source = __import__("json").loads(text) except Exception: return {} if not isinstance(source, dict): return {} metadata = source.get("usageMetadata") if not isinstance(metadata, dict): # 展开包装: {"response": {...}} response = source.get("response") if isinstance(response, dict): metadata = response.get("usageMetadata") if not isinstance(metadata, dict): return {} def _first_int(*keys: str) -> int: for key in keys: value = metadata.get(key) if value is not None: try: return int(value) except (TypeError, ValueError): continue return 0 prompt_tokens = _first_int("promptTokenCount", "total_input_tokens", "input_tokens") reasoning_tokens = _first_int("thoughtsTokenCount", "total_thought_tokens") candidates_tokens = _first_int("candidatesTokenCount", "total_output_tokens") completion_tokens = candidates_tokens + reasoning_tokens total_tokens = _first_int("totalTokenCount", "total_tokens") or (prompt_tokens + completion_tokens) # 缓存命中数:旧版 cachedContentTokenCount / 新版 total_cached_tokens,及 Anthropic 风格明细 cached_tokens = _first_int( "cachedContentTokenCount", "cached_content_token_count", "total_cached_tokens", "cached_tokens", "cache_read_input_tokens", ) if not cached_tokens: for details_key in ("prompt_tokens_details", "input_tokens_details"): details = metadata.get(details_key) if isinstance(details, dict): value = details.get("cached_tokens") if value is not None: try: cached_tokens = int(value) except (TypeError, ValueError): cached_tokens = 0 break return { "prompt_tokens": prompt_tokens, "completion_tokens": completion_tokens, "total_tokens": total_tokens, "cached_tokens": cached_tokens, } async def record_usage_log(entry: Dict[str, Any]) -> None: """ 写入一条请求日志。任何异常仅记录告警,绝不影响主请求流程。 日志先进内存队列由后台任务批量落库,不阻塞请求热路径。 必要字段: model_name, success 可选字段: api_format, client_ip, request_id, credential_email, is_stream, status_code, use_time, prompt_tokens, completion_tokens, total_tokens, cached_tokens, error_message, request_log_file(请求体详细日志的磁盘文件索引,空=未记录) """ await _write_log(entry) async def clear_usage_logs(days: int) -> int: """清理 N 天前的日志,返回删除条数""" await flush_usage_logs() database = await get_database() cutoff = time.time() - days * 86400 return await database.delete_usage_logs_before(cutoff)