From 9c6958e7cea167b04b4a1d342516a6c607c9b4bd Mon Sep 17 00:00:00 2001
From: ROTl24 <119035617+ROTl24@users.noreply.github.com>
Date: Fri, 11 Sep 2026 02:47:20 +0800
Subject: [PATCH] =?UTF-8?q?perf(csv):=20=E6=89=B9=E9=87=8F=E7=A7=BB?=
 =?UTF-8?q?=E9=99=A4=E7=A9=BA=E8=A1=8C=E4=BB=A5=E6=B6=88=E9=99=A4=E8=A1=A8?=
 =?UTF-8?q?=E6=A0=BC=E8=BD=AC=E6=8D=A2=E7=9A=84=E5=B9=B3=E6=96=B9=E7=BA=A7?=
 =?UTF-8?q?=E5=BC=80=E9=94=80?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

CSV 清理函数逐次 pop(0) 或 pop(1)，每次都移动剩余元素，导致大量前导空行或表头后空行的转换耗时呈平方级增长。

改为先定位表头、数据起点和末尾，再逆序删除空行区间；保留内部空行、表头及行对象，不改变列宽或转义逻辑。新增四个十万空行场景，覆盖前导、表头后、尾部与全空输入，同时检查内部空行和宽行保留行为。

验证：37 项 CSV 测试通过；511 种空行排列与基线函数输出和对象身份一致；pre-commit 全文件检查及 diff 空白检查通过。Windows / Python 3.13.15 的合成输入测量中，十万前导空行的真实转换入口三次中位数从 0.6953 秒降至 0.0084 秒。

未运行整个项目的可选格式、远程服务或付费模型测试。性能测量针对合成空行输入，不代表普通文档的整体提速。
---
 .../markitdown/converters/_csv_converter.py   | 32 +++++++++++------
 .../markitdown/tests/test_csv_blank_runs.py   | 35 +++++++++++++++++++
 2 files changed, 56 insertions(+), 11 deletions(-)
 create mode 100644 packages/markitdown/tests/test_csv_blank_runs.py

diff --git a/packages/markitdown/src/markitdown/converters/_csv_converter.py b/packages/markitdown/src/markitdown/converters/_csv_converter.py
index 82fd29747..1aee0ffb8 100644
--- a/packages/markitdown/src/markitdown/converters/_csv_converter.py
+++ b/packages/markitdown/src/markitdown/converters/_csv_converter.py
@@ -30,17 +30,27 @@ def _escape_table_cell(value: str) -> str:
 
 def _trim_outer_blank_rows(rows: list[list[str]]) -> None:
     """Remove empty rows from the beginning and end, and immediately after the header. This operation is performed in-place."""
-    # Pop empty rows from the beginning
-    while len(rows) > 0 and not rows[0]:
-        rows.pop(0)
-
-    # Pop empty rows after the header
-    while len(rows) > 1 and not rows[1]:
-        rows.pop(1)
-
-    # Pop empty rows from the end
-    while len(rows) > 0 and not rows[-1]:
-        rows.pop(-1)
+    start = 0
+    while start < len(rows) and not rows[start]:
+        start += 1
+
+    if start == len(rows):
+        rows.clear()
+        return
+
+    header_index = start
+    start += 1
+    while start < len(rows) and not rows[start]:
+        start += 1
+
+    end = len(rows)
+    while end > start and not rows[end - 1]:
+        end -= 1
+
+    # Remove each blank run at once, rather than shifting the list per row.
+    del rows[end:]
+    del rows[header_index + 1 : start]
+    del rows[:header_index]
 
 
 class CsvConverter(DocumentConverter):
diff --git a/packages/markitdown/tests/test_csv_blank_runs.py b/packages/markitdown/tests/test_csv_blank_runs.py
new file mode 100644
index 000000000..357af591e
--- /dev/null
+++ b/packages/markitdown/tests/test_csv_blank_runs.py
@@ -0,0 +1,35 @@
+"""CSV conversion preserves table contents when trimming long blank runs."""
+
+import io
+
+import pytest
+
+from markitdown import MarkItDown, StreamInfo
+
+
+@pytest.mark.parametrize("position", ["leading", "after_header", "trailing", "all"])
+def test_csv_long_blank_runs(position: str) -> None:
+    blank = b"\n" * 100_000
+    header = b"name,value\n"
+    # An internal blank row and a wider data row must survive trimming.
+    data = b"Alice,1\n\nBob,2,extra\n"
+    content = {
+        "leading": blank + header + data,
+        "after_header": header + blank + data,
+        "trailing": header + data + blank,
+        "all": blank,
+    }[position]
+
+    result = MarkItDown(enable_plugins=False).convert_stream(
+        io.BytesIO(content),
+        stream_info=StreamInfo(extension=".csv", charset="utf-8"),
+    )
+
+    expected = (
+        "| name | value |  |\n"
+        "| --- | --- | --- |\n"
+        "| Alice | 1 |  |\n"
+        "|  |  |  |\n"
+        "| Bob | 2 | extra |"
+    )
+    assert result.markdown == ("" if position == "all" else expected)