You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Local range download leaves an output file after source-read failure despite delete_on_failure=True #2223
On trunk 8a0ef626, the local driver's range-download method creates a destination file but does not clean it up when opening the source fails, even with delete_on_failure=True.
This reproducer is sequential: obtain the object metadata, explicitly remove the stored file, then try to download. It does not depend on winning a filesystem race.
The public range-download contract describes delete_on_failure as deleting a partially downloaded file when the download was unsuccessful, with hash mismatch/file size given as examples. I expected the newly created empty output to be removed for a source-read failure as well. Otherwise a failed operation leaves a file that can be mistaken for a valid empty download.
The local implementation accepts the flag but never reads it. It opens the destination before consuming the source iterator and has no failure cleanup around that loop.
Checks with unchanged source, Python 3.12/Linux:
Range download with delete_on_failure=True: raises FileNotFoundError, leaves a zero-byte output.
Range download with delete_on_failure=False: same result.
Full download_object with the same missing source: returns False, leaves no output.
Successful range [0, 3) before removing the source: returns True, writes b"pay".
No code change yet. This report is limited to failure cleanup for local range downloads; it does not claim a concurrent-write reproduction or a full test-suite pass.
Activity