mirror of
https://ghfast.top/https://github.com/aeroxw/tick-stock-panel.git
synced 2026-09-12 15:34:16 +08:00
perf(ext-data): CSV 编码转换改为分块进行, 峰值内存不再随文件线性增长
This commit is contained in:
@@ -1,6 +1,7 @@
|
||||
"""扩展数据服务 — 配置管理 + 文件解析 + Parquet 存储。"""
|
||||
from __future__ import annotations
|
||||
|
||||
import codecs
|
||||
import copy
|
||||
import json
|
||||
import logging
|
||||
@@ -453,6 +454,44 @@ def apply_config_mapping(df: pl.DataFrame, config: ExtConfig, data_dir: Path) ->
|
||||
return df
|
||||
|
||||
|
||||
# 编码识别与转换的分块大小,与 ext_data 上传写入用的块大小一致。
|
||||
_TRANSCODE_CHUNK_BYTES = 1024 * 1024
|
||||
|
||||
|
||||
def _decodes_as(file_path: Path, encoding: str) -> bool:
|
||||
"""整个文件能否按 encoding 完整解码,逐块判断,不把文件读进内存。"""
|
||||
decoder = codecs.getincrementaldecoder(encoding)()
|
||||
try:
|
||||
with file_path.open("rb") as src:
|
||||
while chunk := src.read(_TRANSCODE_CHUNK_BYTES):
|
||||
decoder.decode(chunk)
|
||||
decoder.decode(b"", True) # 结尾处的半个字符也算解码失败
|
||||
except UnicodeDecodeError:
|
||||
return False
|
||||
return True
|
||||
|
||||
|
||||
def _transcode_to_utf8(file_path: Path, out_path: Path, encoding: str) -> bool:
|
||||
"""按 encoding 逐块转成 UTF-8 写入 out_path;解码失败则删除半成品返回 False。
|
||||
|
||||
增量解码器负责跨块边界的多字节字符:GBK 一个汉字两字节,正好落在块边界
|
||||
上时前半截会被留到下一块,不会被误判成解码失败。
|
||||
"""
|
||||
decoder = codecs.getincrementaldecoder(encoding)()
|
||||
try:
|
||||
with (
|
||||
file_path.open("rb") as src,
|
||||
out_path.open("w", encoding="utf-8", newline="") as dst,
|
||||
):
|
||||
while chunk := src.read(_TRANSCODE_CHUNK_BYTES):
|
||||
dst.write(decoder.decode(chunk))
|
||||
dst.write(decoder.decode(b"", True))
|
||||
except UnicodeDecodeError:
|
||||
out_path.unlink(missing_ok=True)
|
||||
return False
|
||||
return True
|
||||
|
||||
|
||||
def ensure_utf8_csv(file_path: Path) -> Path:
|
||||
"""确保 CSV 文件以 UTF-8 编码可读,非 UTF-8(如 GBK/GB18030)则转换。
|
||||
|
||||
@@ -463,27 +502,14 @@ def ensure_utf8_csv(file_path: Path) -> Path:
|
||||
返回值:若已是 UTF-8 则返回原路径;否则在同目录写一个 *.utf8 文件并返回它
|
||||
(调用方用临时目录,随目录一起清理)。
|
||||
"""
|
||||
raw = file_path.read_bytes()
|
||||
# BOM 处理:UTF-8-SIG 等带 BOM 文件直接交给 Polars(它认识 BOM)
|
||||
try:
|
||||
raw.decode("utf-8")
|
||||
if _decodes_as(file_path, "utf-8"):
|
||||
return file_path # 已是合法 UTF-8
|
||||
except UnicodeDecodeError:
|
||||
pass
|
||||
# 依次尝试常见中文编码,第一个能完整解码的即为命中
|
||||
for enc in ("gb18030", "gbk", "gb2312", "big5"):
|
||||
try:
|
||||
text = raw.decode(enc)
|
||||
except UnicodeDecodeError:
|
||||
continue
|
||||
out_path = file_path.with_suffix(file_path.suffix + ".utf8")
|
||||
# newline="" 关闭写入时的换行转换。默认转换在 Windows 上把文本里的 \n
|
||||
# 写成 \r\n,源文件本来就是 CRLF 时就变成 \r\r\n,多出来的 \r 被 Polars
|
||||
# 当作最后一列内容的一部分:列名变成 "收盘价\r",每行的值变成 "12.34\r",
|
||||
# 该列于是被推断为字符串而不是数值。本函数针对的同花顺/东财/通达信和
|
||||
# Windows Excel 导出文件用的正是 CRLF。
|
||||
with out_path.open("w", encoding="utf-8", newline="") as f:
|
||||
f.write(text)
|
||||
if not _transcode_to_utf8(file_path, out_path, enc):
|
||||
continue
|
||||
logger.info("CSV 编码转换 %s → %s (%s)", file_path.name, out_path.name, enc)
|
||||
return out_path
|
||||
# 都无法解码:返回原路径,让 Polars 抛出更精确的原始错误
|
||||
|
||||
@@ -0,0 +1,72 @@
|
||||
"""CSV 编码转换的内存占用回归测试。
|
||||
|
||||
上传路径专门用 `_write_upload_capped` 分块落盘(issue #204),docstring 写明
|
||||
是为了「避免 `await file.read()` 把整个文件读入内存(大文件可能触发高内存占用、
|
||||
进程 OOM 或服务不可用)」。但紧接着的 `ensure_utf8_csv` 曾用 `read_bytes()`
|
||||
把同一个文件整个读回内存再整体解码,把这层保护抵消掉:50MB 上限的文件实测
|
||||
峰值 302MB(6.05×),解码出的 str 比原字节还大。
|
||||
|
||||
这里用 tracemalloc 量峰值,判据是「与文件大小无关」而不是某个绝对值:转换
|
||||
按块进行时峰值只跟块大小有关,文件翻倍不会让峰值翻倍。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import tracemalloc
|
||||
from pathlib import Path
|
||||
|
||||
from app.services.ext_data import ensure_utf8_csv
|
||||
|
||||
_ROW = "浦发银行,600000,12.34,上海证券交易所\r\n"
|
||||
_HEADER = "名称,代码,收盘价,交易所\r\n"
|
||||
|
||||
|
||||
def _gbk_csv(path: Path, size_bytes: int) -> Path:
|
||||
body = _ROW * (size_bytes // len(_ROW.encode("gb18030")))
|
||||
path.write_bytes((_HEADER + body).encode("gb18030"))
|
||||
return path
|
||||
|
||||
|
||||
def _peak_bytes(path: Path) -> int:
|
||||
tracemalloc.start()
|
||||
try:
|
||||
ensure_utf8_csv(path)
|
||||
_, peak = tracemalloc.get_traced_memory()
|
||||
finally:
|
||||
tracemalloc.stop()
|
||||
return peak
|
||||
|
||||
|
||||
def test_transcode_peak_memory_is_far_below_the_file(tmp_path: Path) -> None:
|
||||
path = _gbk_csv(tmp_path / "big.csv", 8 * 1024 * 1024)
|
||||
|
||||
peak = _peak_bytes(path)
|
||||
|
||||
# 分块转换时峰值只跟块大小有关。修复前是文件的 6 倍。
|
||||
assert peak < path.stat().st_size
|
||||
|
||||
|
||||
def test_transcode_peak_memory_does_not_grow_with_the_file(tmp_path: Path) -> None:
|
||||
small = _peak_bytes(_gbk_csv(tmp_path / "small.csv", 2 * 1024 * 1024))
|
||||
large = _peak_bytes(_gbk_csv(tmp_path / "large.csv", 8 * 1024 * 1024))
|
||||
|
||||
# 文件大 4 倍,峰值不应跟着涨:留一倍余量给解释器噪声。
|
||||
assert large < small * 2
|
||||
|
||||
|
||||
def test_multibyte_character_on_a_chunk_boundary_survives(tmp_path: Path) -> None:
|
||||
# GBK 一个汉字两字节,分块时可能正好被切开。增量解码器负责把半个字符
|
||||
# 留到下一块;若改成逐块独立 decode,这里会解码失败并整份回退。
|
||||
import polars as pl
|
||||
|
||||
from app.services.ext_data import _TRANSCODE_CHUNK_BYTES
|
||||
|
||||
filler = "浦发银行" * ((_TRANSCODE_CHUNK_BYTES // 8) + 1)
|
||||
text = f"名称,备注\r\n浦发银行,{filler}\r\n"
|
||||
path = tmp_path / "boundary.csv"
|
||||
path.write_bytes(text.encode("gb18030"))
|
||||
assert path.stat().st_size > _TRANSCODE_CHUNK_BYTES
|
||||
|
||||
df = pl.read_csv(ensure_utf8_csv(path), infer_schema_length=10000)
|
||||
|
||||
assert df.columns == ["名称", "备注"]
|
||||
assert df["备注"][0] == filler
|
||||
Reference in New Issue
Block a user