Crashing during snapshot deletion might result in unreferenced data left in repository

[nkvoll]

It's currently possible to end up with unreferenced data in a snapshot repository, given the following steps:

1. Create a index, foobar, with size X bytes
2. Snapshot cluster
3. Start deleting the snapshot, crash after deleting snapshot-{} and metadata-{} files
4. Delete index foobar
5. Snapshot cluster again

Normally, step 5 would cause the files no longer referenced by any snapshots do be deleted, but if the underlying index is deleted as well, they won't get cleaned up. In the example above, there would be X bytes of disk space used without any snapshot referencing them. Given sufficiently large values of X, this could be a significant amount of storage wasted. Even with small amounts of data, this might accrue over time to become significant.

Suggestion: create a deleting-{} file as a sibling of the snapshot-{} file that gets written before the files referenced by the snapshot gets deleted. When the deletion has been completed, this file should be the last one deleted. These files indicates that a deletion is in progress or have been attempted so it's possible to tell that the snapshot might be in a half-deleted state (so we can avoid using it). It should also enable later snapshot processes to continue the deletion process where the previous left off.

[tlrx]

We talked about it today with @ywelsch and we still think that this is a real issue. We don't have any plan about it yet but this is something we need to think about for the future of the snapshot/restore functionality.

[original-brownbear]

There is another scenario here that can lead to the same stale data that I found:

1. Create a index, foobar, with size X bytes
2. Start snapshotting cluster
3. Abort snapshot
   4 .Temporarily disconnect datanode from master (via network partition or so) while it is writing shard data to repository
4. Master node finishes deleting snapshot

If in this scenario the data node continues to write segment files (as seen in this test failure #39852) we get unreferenced data since step 5 has removed any metadata reference to the index and snapshotting it again will lead to a new index uuid.

[original-brownbear]

I opened #40228 with a suggestion on how to write the tombstones to get a better handle on this situation.