Hi Nathan,
The process on L is indeed (or should I say "remove_indeed") has a very different flow. It looks to me that this behavior was most likely a regression relative to L, though.
On L, the logic is unrolled in RGWLC::bucket_lc_process, into two large cases: !versioned bucket, and else (the case has a misleading comment "/* bucket versioning is enabled or suspended */"). So if there can be a delete marker, it's only in the else case, and, in the else case, remove_expired_obj can be reached only if we dont take the branch (my l.555):
if (obj_iter->is_delete_marker()) {
if ((obj_iter + 1)==objs.end()) {
if (is_truncated) {
//deal with it in next round because we can't judge whether this marker is the only version
list_op.get_next_marker() = obj_iter->key;
break;
}
} else if (obj_iter->key.name.compare((obj_iter + 1)->key.name) == 0) { //*obj_iter is delete marker and isn't the only version, do nothing.
continue;
}
skip_expiration = prefix_iter->second.dm_expiration;
remove_indeed = true; //we should remove the delete marker if it's the only version
} else {
remove_indeed = false;
}
Which looks like the same logic from the upstream PR l. 729?
}
if (o.is_delete_marker()) {
if (oc.ol.next_has_same_name()) {
return false;
} else {
*exp_time = real_clock::now();
return true;
}
}
Matt