Tree changes in query
* Move inner selection S to a temporary table
* Reorder join field in outer query
* Remove intersection with File view in outer selection
File view is already intersected with in selection S
With the above-mentioned changes, on a specific testset, query took 9s instead of 6min.
The index on Block.Size improves the performance of the query that
checks that the user has not altered the block size of the backup
configuration.
The index on Block.Hash and Block.VolumeID improves the performance of
the query that obtains the blocklists.
There are anecdotes that these indices can drastically improve
performance (queries that used to take hours completed in seconds with
the index):
https://forum.duplicati.com/t/repairing-a-big-database/8174/4https://github.com/duplicati/duplicati/issues/3884#issuecomment-612296016
This fixes issue #3884.
When a backup is interrupted it will now show as a partial backup in the list of backups on the restore files page.
Temporary filelist volumes are protected from deletion in several places. All but the RestoreHandler
only protect the most recent filelist. The RestoreHandler protects all temporary filelist volumes as the user
might be restoring from one that would be deleted.
Prevent the CoCoL BackendRequests channel is retired exception from appearing when the backup is stopped by checking
if the channel is retired before trying to add the FlushRequest to the channel. Any exceptions from the BackendUploader
task will still be thrown.
The exception could occur due to the existing delete for FilesetEntry
deleting based on volume ids, which would only match a filelist.
The bsIdsSubQuery would return BlocksetIds that should
be deleted and would delete from the FileLookup table.
FilesetEntry would then have a higher count of rows
than what would be returned by the LEFT OUTER JOINS on all of the
other tables in VerifyConsistency() since more rows had been deleted
from FileLookup.
To fix this when rows are deleted in FileLookup, use the associated metadata id
to delete from FilesetEntry and other tables based on the blockset id for the
metadata id.
Some existing delete statements were moved, none were deleted.
The existing FileLookup delete had the match on metadata id removed as
a blockset id does not equal a metadata id.
One of the first things UploadSyntheticFilelist does is throw an exception if the volume is not in the Uploaded state. Or if it is in the Uploaded state, the next check would prevent a synthetic filelist from being created if the volume was not in the Uploading or Temporary states.
Removed the check for the volume not being in the Uploaded state as the warning in the second check handles that case.
Removed the using statement as it would dispose of the FilesetVolumeWriter after passing it to the UploadChannel. The UploadChannel could not use it since it had been disposed.
Fixed the GetRemoteVolumeFromIDAsync taking a fileset id and matching it up against a volume id. Now it will use fileset id to get the volume that matches the fileset.
Change the StopNow disruption test to not use a synthetic file list. The test was written with it not working and therefore fails a check on the number of filesets after stopping a backup.
Bonus fix for a possible null exception in Dispose for VolumeWriterBase.
The LocalDatabase class was deleting from the BlocklistHash table based on the volume id of blocks. The deletes for the Blockset and BlocksetEntry tables were using a union query which included the same query the BlocklistHash delete was using. Changed the BlocklistHash delete query to use the union as well to ensure that all three tables would use the same query to delete the necessary rows. This fixes the "Detected non-empty blocksets with no associated blocks!" error.
Moved a write call in the DataBlockProcessor to send a volume for upload after setting blockvolume to null. This prevents a null reference exception because otherwise it would be sent to the SpillCollector and it tries to add a block, but the volume has already been closed.
Added the ? operator to the RegistryUtility class to prevent a null reference exception while debugging. It was annoying catching it in the debugger.
Fixes#4037
The ROW_NUMBER window function is not available in SQLite versions prior
to 3.25. In some cases, I think we do fall back to the system SQLite
libraries, and many distributions do not have versions that support the
window functions. To be safe, we will generate the backup versions
manually.
The ID in the Fileset table is the auto-incremented integer. When
considering filesets for deletion (in particular, when deleting specific
versions), we need to reference the backup version number that is
exposed to the user (integers where the newest backup version is 0).
This extracts the logic for various fileset deletion strategies into
separate classes. This simplifies the code and makes it easier to test
the individual functionalities.
Honoré Vasconcelos