List every version
--all-versions lists every S3 version and delete marker, every GCS generation, or every Azure Blob version, snapshot and soft-deleted blob, instead of the current objects only.
swath list --all-versions s3://BUCKET[/PREFIX] -o DIR # every version and delete marker (ListObjectVersions)
swath list --all-versions gs://BUCKET[/PREFIX] -o DIR # every generation, live and noncurrent (versions=true)
swath list --all-versions az://ACCOUNT/CONTAINER[/PREFIX] -o DIR # versions, snapshots, soft-deleted blobsRows. One row per S3 version or delete marker, per GCS generation, or per Azure Blob entry (a version, a snapshot, a soft-deleted blob, or the current blob), under the version schemas. A delete marker's absent metadata is null, never 0. version_id is verbatim: "null" is a real ID (an object written while versioning was off). The current objects are is_latest AND NOT is_delete_marker (S3) or is_latest (GCS and Azure Blob); a listing of the current objects alone is the default mode, which is cheaper.
Order. Keys ascend; a key's rows are consecutive, in the service's order within the key: S3 newest first, versions and delete markers interleaved; GCS by generation, ascending; Azure Blob as the service lists a name's entries, keyed by their time within the name: snapshots and versions oldest first, the current blob last. S3 can list two entries of a key with the same last_modified (to the second), so that time does not rebuild S3's order: the file order does. In an unfiltered, complete listing, the first row of an S3 key's group is the one with is_latest, and a delete marker there means the object is deleted. On GCS, the last row of a name's group is the newest generation; it is live (is_latest) only if deleted_at is null, and a GCS name may have no live generation. On Azure Blob a name's is_latest row, when it has one, is its last; a versioned blob that was deleted has only previous versions, and a soft-deleted one only is_deleted rows. A filter can remove a key's latest version, so in a filtered dataset read is_latest per row instead of by position.
Parts never split a key. A part's size target, --part-max-rows and --part-interval cut at the next key, so a part can pass them by one key's history; parts stay key-disjoint, and a resume never lists half a key again. The interval does not bound how long a part stays open while one key's rows are still arriving.
Parallelism. Ranges split only between keys, so one key with a very long history is listed by one worker, page after page, while the rest of the bucket proceeds. S3's versions call is slower than ListObjectsV2: about 1.85 times the time per request at the median of our measurements, and pages of old versions are slow to first byte; history multiplies the pages besides. GCS lists generations at the cost of live objects. Azure Blob lists every entry with the same call and the same format (Arrow where the account has it): each request costs a listing operation, as always, and the entries multiply the pages; Microsoft recommends fewer than 1,000 versions per blob for listing performance.
Permissions and refusals. S3 needs s3:ListBucketVersions (s3:ListBucket alone answers 403, and swath never falls back to a current-object listing); GCS needs the same storage.objects.list; Azure Blob the same permission to list (a SAS with l, anonymous access to a public container). Text rows carry the versions schema's columns: S3's version_id, is_latest and is_delete_marker (a delete marker's size and etag null), GCS's generation, is_latest and deleted_at (null on the live generation), Azure Blob's version_id, is_latest, snapshot, is_deleted, deleted_at and remaining_retention_days; the table shows the version ID (empty for an Azure snapshot), latest and, on S3, delete_marker. --min-size and --max-size never match a delete marker. A resume continues in the mode its run started in; a run in the other mode is refused on that output. Directory buckets have no versions and stay refused.
The run report's version counts (under output) describe this attempt after filtering: S3 versions and delete_markers sum to emitted rows; GCS and Azure Blob versions include the noncurrent rows (on Azure every row that is not a name's latest: previous versions, snapshots, soft-deleted blobs). distinct_keys counts surviving names, and live_objects counts live versions among those rows. Table output includes the version ID and latest state, and for S3 the delete-marker state, in wire order.