Rename into place

How it works

Rename on the same filesystem is the signal. A copy across filesystems is not atomic. Object stores emulate this with a complete call, not a rename. The idea is the same: do not publish early.

A job writes feed.tmp, finishes, and renames to feed.csv. The importer only watches feed.csv. A crash leaves feed.tmp, which nobody imports. Last month a direct write to feed.csv was imported at 10 percent.

How it differs

Rename is not versioning. It is a publish step.

Cross-device rename copies and can leave the temp name.

Write temp on the destination volume.

On the ticket

• The practical close is a log line: time, actor, byte count, result.
• Without that line the transfer is a story.
• With it, the next person can see whether this door did what the ticket claimed.
• If the path is shared, say so in the partner profile so a later change does not silently pick a different limit, key, or region.
• On a real ticket, write down the door, the byte count, and the clock.
• For rename into place, that means naming the host or bucket, the expected size, and the time the other side must have a complete file.
• A progress bar is not that record.
• A 200 response that arrives before the complete call is not that record.
• If a retry is allowed, say how many and whether it resumes.
• If a person must approve the send, name the person.
• Partners who receive rename into place files should match on hash or size before they import.
• A same-length corrupt file passes a size check and fails a hash.
• Keep the published hash off the only channel an attacker can edit, or treat it as a corruption check rather than a substitution check.
• When the path changes, new key, new region, new cap, update the profile the same day so the next run does not use a stale limit.