Context: Selective point-in-time backup/restore follows this high-level flow:

1. Backup:
   1. Create a consistent snapshot at checkpoint timestamp Tbackup.
   2. Copy the data files for the selected collections at Tbackup.
   3. Capture the entire oplog starting at Tbackup (non-selective).
2. Restore at point in time Trestore.
   1. Restore the data files for the selected collections (at Tbackup state).
   2. Replay all the oplog from Tbackup to Trestore (non-selective).

Both oplog capture and replay are non-selective, and instead rely on trying to apply the oplog entry and if it fails to apply because the namespace doesn't exist then we ignore the error. (see also: SERVER-62615).

The issue: Since both (1.3) and (2.2) are non selective, any new collections or views are created after Tbackup end up in the final restored cluster. For example consider:

• A replica set has collections a, b, c and d.
• The user selects to only backup collections a and b.
• After Tbackup, a new collection e is created.
• The user starts restoring the backup. This initially creates collections a and b.
• During application of the oplog, entries for collections c and d become no-ops.
• However the oplog entry to create new collection e is still applied successfully.
• As a result the restored replica set has collections a, b and e (despite the user selecting only a and b).

It seems that during selective restore we should skip the oplog entries to create new collections/views.