src/data/loader.py

"""
L2 数据加载器（基于 HF 数据集，按需下载单日文件）。

数据集: kangkangchen/a-share-l2-600809
         kangkangchen/a-share-tushare-context-600809
"""

from __future__ import annotations

import os
from typing import Dict, Iterable, List

import pandas as pd
from huggingface_hub import hf_hub_download

L2_REPO = "kangkangchen/a-share-l2-600809"
TUSHARE_REPO = "kangkangchen/a-share-tushare-context-600809"
TS_CODE = "600809.SH"

# 2025-09-25: 上交所被动方 order-link 命中率异常，整日剔除
BLACKLIST_DATES = {20230102, 20250925}

L2_TABLES = ("level2_trades", "level2_orders")


def _parquet_path(table: str, date: int) -> str:
    ds = str(date)
    year, month = ds[:4], ds[4:6]
    if table in ("level2_quotes", "level2_orders", "level2_trades"):
        return f"data/{table}/year={year}/month={month}/ts_code={TS_CODE}/{ds}.parquet"
    # tushare tables
    if table == "moneyflow_hsgt_daily":
        return f"data/{table}/year={year}/month={month}/*.parquet"  # market-level, no ts_code
    return f"data/{table}/year={year}/month={month}/ts_code={TS_CODE}/*.parquet"


def download_l2_table(table: str, date: int) -> str:
    """下载单日单张 L2 表，返回本地 parquet 路径（使用 HF 缓存）。"""
    if date in BLACKLIST_DATES:
        raise ValueError(f"date {date} is blacklisted")
    path = _parquet_path(table, date)
    return hf_hub_download(L2_REPO, path, repo_type="dataset")


def load_trades(date: int) -> pd.DataFrame:
    """加载单日逐笔成交。"""
    p = download_l2_table("level2_trades", date)
    return pd.read_parquet(p)


def load_orders(date: int) -> pd.DataFrame:
    """加载单日逐笔委托。"""
    p = download_l2_table("level2_orders", date)
    return pd.read_parquet(p)


def load_l2_day(date: int) -> Dict[str, pd.DataFrame]:
    """加载单日 trades + orders。"""
    return {
        "trades": load_trades(date),
        "orders": load_orders(date),
    }


def list_available_dates(
    start: int = 20230101, end: int = 20260331
) -> List[int]:
    """
    通过 HF API 列出可用日期。
    注意：这需要一次 API 调用。在 sandbox 里可以直接扫描快照目录。
    本地测试时手动指定日期范围即可。
    """
    from huggingface_hub import list_repo_tree

    dates = set()
    for table in ("level2_quotes", "level2_orders", "level2_trades"):
        tree = list_repo_tree(
            L2_REPO, repo_type="dataset", recursive=True
        )
        for item in tree:
            path = item.path
            if path.endswith(".parquet") and "ts_code=600809.SH" in path:
                try:
                    d = int(os.path.splitext(os.path.basename(path))[0])
                    if start <= d <= end and d not in BLACKLIST_DATES:
                        dates.add(d)
                except ValueError:
                    pass
    return sorted(dates)


def load_tushare_table(root: str, table: str) -> pd.DataFrame:
    """加载 tushare 单表全量。root 是 HF snapshot 本地路径。"""
    import glob

    if table == "moneyflow_hsgt_daily":
        pattern = f"{root}/data/{table}/year=*/month=*/*.parquet"
    else:
        pattern = f"{root}/data/{table}/year=*/month=*/ts_code={TS_CODE}/*.parquet"

    files = sorted(glob.glob(pattern))
    if not files:
        return pd.DataFrame()
    dfs = [pd.read_parquet(f) for f in files]
    df = pd.concat(dfs, ignore_index=True)
    if "trade_date" in df.columns:
        df = df.sort_values("trade_date").reset_index(drop=True)
    return df