Repository navigation
Description
Tip
Help move this idea forward
- Use the 👍 reaction to show support for this feature.
- Avoid commenting unless you have relevant information to add; unnecessary comments create noise for subscribers.
- Subscribe to receive notifications about status changes and new comments.
Is your feature request related to a problem? Please describe.
Public bodies must keep records for their full statutory retention period, and may only destroy them through a recorded decision. A record under legal hold may not be destroyed at all. The natural setup for this is Nextcloud on top of storage that enforces retention, such as an S3 bucket with Object Lock. Today the two cannot talk to each other:
- Deletion is an unconditional instruction. A backend has no way to answer "not yet, and here is why".
- On S3 it is worse than an error.
AmazonS3::unlink()(files_external) andS3ObjectTrait::deleteObject()(primary object storage) sendDeleteObjectwithout aVersionId. Object Lock requires versioning, so S3 answers with a delete marker and reports success. The file disappears for the user and from Nextcloud, while the locked bytes stay in the bucket out of sight. Folder deletes usedeleteObjectswithQuiet: trueand do not read back the per-key errors. - When a deletion is refused today (a listener throws
AbortedEventException, or the storage returnsfalse), WebDAV answers a bare 403 and the reason is lost.
We found this by reading the code on master; we are confirming it on MinIO with Object Lock and will add the results here.
Describe the solution you'd like
- A typed refusal. A storage or listener can throw a refusal carrying a reason for the user and, optionally, the date until which the object is held. WebDAV returns it with the message, and Files shows it, also when emptying the trash.
- Permanent deletion as a request. Emptying the trash asks the backend, and the backend may refuse or defer. A refused item stays where it is.
- Events for both a refused and a completed deletion, carrying the file's identity, so audit logging can record both.
- S3 storages that respect Object Lock: recognise retention and legal hold and report them as a refusal instead of writing a delete marker, and check a folder before deleting it so it is never left half deleted.
Core keeps no retention rules. Which object is kept for how long belongs to the backend or to apps; core only understands the answer and shows it.
Describe alternatives you've considered
- files_retention. Despite the name it solves the opposite half: it deletes tagged files after a period, and nothing stops a user from deleting them earlier. Records need protection until the end date first, and only then destruction with a recorded decision. files_retention would benefit from this proposal: its background job calls
Node::delete(), so it would get a clear "not yet" back instead of silently creating a delete marker on S3. - App-level protection, such as a listener on
BeforeNodeDeletedEvent. It does not hold for records: an admin can disable the app, other code paths bypass it, and it cannot protect against direct access to the bucket. Only the storage layer can enforce retention; core needs to understand its answer.
Additional context
This is one of a small set of storage interface extensions for records-managed backends, together with backend metadata (#66355) and durable object identity (#66499). We (Conduction, working with Dutch municipalities and VNG Common Ground) would like to build and maintain the pull requests, with tests and a MinIO Object Lock test setup.
Metadata
Metadata
Assignees
Labels
Type
Projects
- StatusShow more project fieldsTo triage
Activity
[-]Durable object identity that survives a cache rebuild[/-][+]let a storage backend refuse or defer a deletion, with a reason shown to the user[/+][-]let a storage backend refuse or defer a deletion, with a reason shown to the user[/-][+]Let a storage backend refuse or defer a deletion, with a reason shown to the user[/+]rubenvdlinde commentedon Nov 17, 2026
We ran the scenarios from this issue on a test setup to replace "found by reading the code" with observed behaviour.
Setup: Nextcloud 34.0.0.12 (SQLite), MinIO RELEASE.2025-04-22, primary object storage on a bucket with Object Lock (governance, 1 day), plus files_external S3 mounts on a locked bucket, a plain bucket, and a plain bucket with a policy that denies
s3:DeleteObject. The delete paths (AmazonS3::unlink/rmdir/batchDelete,S3ObjectTrait::deleteObject,ObjectStoreStorage::rmObject/rmObjects) are byte-identical on master as of today, so we expect the same behaviour there.Nothing was logged in the locked-bucket cases, also at debug level. After every "delete" the original bytes could be read back by version id.
What this means:
VersionIdis allowed to write a delete marker.batchDelete()sendsdeleteObjectswithQuiet: trueand does not read theErrorsarray. When the storage refuses (in our test a bucket policy denying deletes), a folder delete returns 204 while every object stays, and after the rescan the files have new ids, so shares and tags on them are lost. We saw that with the trashbin both enabled and disabled.We'd like to propose a fix for both, independent of where retention decisions are made: read and report per-key errors from
deleteObjects, and on buckets with Object Lock check retention and legal hold before deleting, so Nextcloud reports a refusal instead of success. Would a PR along those lines be welcome, and is there a preferred way to surface a storage refusal to the user?