[avallete]

TL;DR

Follow-ups to the non-blocking review notes on #329 (merged). A native write now also clears the stale sidecar copy, the sweep resumes inside a large shard instead of restarting it, the sweep lock is only released by its owner, and two comments and REPORT.md statements are corrected.

Before

flowchart LR
  A["Sweep cut short<br/>inside shard ab"] --> B["Cursor = ab"] --> C["Next wake restarts ab<br/>from its first entry"]

Loading

After

flowchart LR
  A["Sweep cut short<br/>inside shard ab"] --> B["Cursor = ab/last swept entry"] --> C["Next wake resumes<br/>after that entry"]

Loading

Why

These address jgoux's review comments on #329:

• A multipart part rewritten in place from an engine with xattrs kept its old sidecar ETag. Switching back to a no-xattr engine then failed completion with InvalidChecksum.
• A pass repeatedly cut short inside one shard restarted that shard on every wake and never reached later shards.
• Two processes taking over one stale lock could delete each other's lock in finally.
• Comments and REPORT.md claimed every path is fresh, but the Iceberg S3 PutObject route is unversioned. They also misdescribed how a relative storage root behaves.

What changed

• setAttributeSync: after a successful native set it drops the sidecar copy of that attribute, sharing one helper with removeAttributeSync.
• Sweep cursor: it records <shard>/<last swept entry> at every batch pause and resumes after that entry, walking each shard in sorted order.
• Sweep lock: it holds a random token, and a process only releases the lock if the token is still its own. A takeover race stays documented as best-effort; it only duplicates removals that are harmless to repeat.
• Comments and REPORT.md: they now name the unversioned Iceberg path and its narrow race with the sweep, and state that a relative root is resolved against the working directory, so the original ENOTSUP remains.
• Sweep lock refresh: the lock is refreshed after each pause too, so a slow batch can't let it go stale while the sweeper is live.
• Unit tests: new cases for the native rewrite dropping the stale copy, resuming inside a shard, and writing the cursor at a pause while keeping another process's lock (20 tests).
• REPORT.md: records the remaining limitation: a multipart part rewritten in place under Docker Desktop keeps its stale native etag on an engine with xattrs, which takes two engine switches within one upload.

Reviewer notes

• Validation-only release runs passed on e668239 (same tree as the head before the test commit), including the delete-then-sweep smoke step on both Linux targets: v1.79.27, v1.79.21. The head (7d4a761) also passed a v1.79.27 run.
• Cost: a native set now reads the sidecar file too. That's normally an immediate ENOENT, so only one extra syscall per attribute write where xattrs work.

Part of supabase/cli#6900

🤖 Generated with Claude Code