The current implementation of Cinder Backup functionality only supports full backup and restore of a given volume. There is no provision to backup changes only that happened since last backup. As the volumes grow bigger and over, all changes to volumes between backups stay relatively small, copying entire volumes during backups will be resource intensive and do not scale well for larger deployments. This specification discusses implementation of incremental backup feature in detail.
Cinder backup API, by default uses Swift as its backend. When a volume is backed up to Swift, Swift creates a manifest file that describes the contents of the backup volume. The manifest file contains header (metadata) and array of pointers to the volume backup files. Since Swift has an upper limit on the object size, Cinder backup API splits the volume data into individual chunks of Swift object size and uploads these individual chunks to Swift. Cinder volume backup manifest file includes these list of objects, their corresponding objects, the logical offset of each object within the volume and a message digest of each chunk to detect any unwarranted changes to objects. During restore operation, Cinder reconstructs the volume based on the manifest and individual chunks referenced in the manifest.
To support incremental backup functionality, we introduce another object called shafile to the list of backup files. shafile file helps track changes to the volume since last backup. This new object holds SHA256s of the volume. A brief description of SHA-2 or SHA256 can be found at http://en.wikipedia.org/wiki/SHA-2. The backup manifest file will have a reference to this object. During a full backup operation, Cinder divides up the volume into fixed blocks of user configurable block size. It calculates SHA256 of each block and compiles a list of SHAs and uploads the shafile to the backup container.
To keep the incremental backup implementation simple, an incremental operation is only performed with respect to a full backup. During incremental backup, cinder reads the shafile of the full backup. It creates a new shafile from the current volume data and compares the new shafile with full backup shafile to calculate the blocks that are changed since last full backup.
We will use existing manifest mechanism to capture the delta. Since full backups do not contain any holes, offset+lengths of each chunk of the volume describe the full length of the volume logical address. However with incremental backup, this model is challenged and the offset/chunk of individual files become sparse. The absence of offset/length in a manifest represents the data that is not modified since last backup. One potential drawback of this approach is if changes to volume are fragmented, incremental backup may result in too many objects in Swift. However object stores like Swift are built to handle many small objects effectively.
The new shafile is uploaded as part of the incremental backup.
The manifest header identifies this backup as incremental backup and hence contains a reference to the full backup container.
Following changes are made to the manifest header of the backup
metadata['version'] = self.DRIVER_VERSION metadata['backup_id'] = backup['id'] metadata['volume_id'] = volume_id metadata['backup_name'] = backup['display_name'] metadata['backup_description'] = backup['display_description'] metadata['created_at'] = str(backup['created_at']) # Changes to metadata section of manifest metadata['shafile'] = <shafilename> # Path to shafile name. Or # can be hardcoded to "shafile" # in the container metadata['backup_type'] = "incrementa/full" # backup type metadata['full_container'] = <object path> # path of full backup
Restore API is not expected to change, however restore implementation will be changed to handle incremental backup. To keep the restore from incremental backup simple and easy to test, the restore operation first performs restore of the full volume from the full backup copy and then apply incremental changes at offset and length as described in the incremental backup manifest.
Snapshot based backups:
Since existing backup implementation copies the data directly from the volume, it requires the volume to be detached from the instance. For most cloud workloads this may be sufficient but other workloads that cannot tolerate prolonged downtimes, a snapshot based backup solution can be a viable alternative. Snapshot based backup will perform a point in time copy of the volume and backup the data from point in time copy. This approach does not require volume to be detached from the instance. Rest of the backup and restore functionality remain the same. As an alternative, snapshot based backup can be implemented by extending existing backup functionality to snapshot volumes. This approach can be lot more simpler than backup API taking snapshot of the volume and then managing the snapshots.
The first benefit can be achieved as a post processing of the backup images to remove duplication or by using dedupe enabled backup storage. However the second benefit cannot be achieved unless Cinder backup supports incremental backup.
No percieved data model changes
No new APIs are proposed. Instead existing backup API will be enhanced to accept additional option called “–incr” with <path to full backup container>” as its argument.
cinder backup-create <volumeid> --incr <full backup container> Performs incremental backup cinder backup-create <volumeid> --snapshot Optionally backup-create will backup a snapshot of the volume. Snapshot based backups can be performed while the volume is still attached to the instance. cinder backup-create <volumeid> --snapshot --incr <full backup container> Optionally backup-create will perform incremental backup from volume snapshot
No anticipated changes to restore api
python-cinderclient will be modified to accept “–incr” option. It may include some validation code to validate if the full backup container path is valid
Currenly backup functionality is not integrated with OpenStack dashboard. When it happens, the dashboard will provide an option for user to choose incremental backup
Except for calculating SHAs during full backup operation, there is no other performance impact on existing API. The performance penalty can be easily offset by the efficiency gained by incremental backup. Also new hardware support CPU instructions to calculate SHAs which alleviates some stress on the CPU cycles.
Primary assignee: muralibalcha(email@example.com)
Other contributors: giribasava(firstname.lastname@example.org)
Unit tests will be added for incremental backup.
- Changes to first block
- Changes to last block
- Changes to odd number of successive blocks
- Changes to even number of successive blocks
- Changes spread across multiple sections of the volume
Need to document new option in the block storage manual.