Optimize makedirs on Windows
@mdboom 已經在處理了。
開始於 2023年2月28日。
- 主要語言
- Python
- 星號
- 77.2k
- 分支
- 36k
- 平均合併
- 1 天 9 小時
- 30 天內合併 PR
- 558
描述
I've been working to improve the performance of some Python workflows on Windows, and I found that a noticeable amount of time was being spent in makedirs.
I some more profiling on the venv create scenario, and learned that roughly 1/3rd of the stat cost can be solely attributed to the exists check in makedirs (src). (Which is being called by compile here, and from my cursory glance at that code, it should simply be passing exist_ok=True rather than catching FileExistsError, right?). What makes this especially wasteful is that this run contained 1451 calls to makedirs and 1455 calls to makedir, implying that the exists check was making things worse, rather than helping. Put another way, exists is taking nearly as much time as mkdir here.
While it's true that if/when we get GetFileInformationByPath(), it should make stat calls much faster, that will only benefit the minority of users running a new version of Windows containing this new API. Therefore, it makes sense for us to improve the state of things for users running an older Windows, when practical.
And in this case, we should allow makedirs on Windows to execute in an idiomatically performant way. And that is to simply call CreateDirectoryW() and selectively ignore ERROR_ALREADY_EXISTS.
CreateDirectoryW() isn't cheap, but in the case of an existing directory, I don't think it's any worse than Windows' stat implementation that uses CreateFileW(). And in the case of when the directory doesn't exist and needs to be created, it will be significantly faster to make the single call.
It's probably worth splitting off a Windows-specific implementation of makedirs into the native side (posixmodule.c ?) so that we don't depend on raising a relatively expensive FileNotFoundError.
The non-Windows implementation should still use the cheap and reliable exists() and isdir().
(Forked from discussion on #101196)
貢獻指南
從這裡開始
- 先讀完整個 Issue,再讀專案的貢獻指南。
- 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
- Fork 儲存庫,在一個分支上完成修改。
- 送出 Pull Request,並在描述裡引用這個 Issue 編號。
評估
這個 Issue 還沒有評估資料。