Alternator: support backup and restore

[nyh]

DynamoDB has two backup features: There's an on-demand backup and restore feature (the user says when to do a backup, and can later restore to that point), and a continuous backup feature, allowing a user to restore the data as it was any time in (near) history.

The API requests we need to implement: For on-demand backup: CreateBackup, DescribeBackup, DeleteBackup, ListBackups, RestoreTableFromBackup; For continuous backup: UpdateContinuousBackups, DescribeContinuousBackups, RestoreTableToPointInTime.

Confusingly, DynamoDB documentation uses the terms "Continuous Backups" and "Point-in-Time Recovery" (PITR) almost interchangeably, but the API (UpdateContinuousBackups, DescribeContinuousBackups) has a separate flag called ContinuousBackupsStatus which is always on and cannot be turned off (so isn't interesting...), and PointInTimeRecoveryStatus is the real flag of whether point-in-time recovery is enabled for this table.

See my discussion of these topics in the relevant sections of https://docs.google.com/document/d/1i4yjF5OSAazAY_-T8CBce9-2ykW4twx_E_Nt2zDoOVs/edit.

In November 2021, DynamoDB also announced a third way of doing DynamoDB backups dubbed AWS Backup (see documentation here). On-demand backups stored in AWS Backup also have an option of being "cold backups" - cold backups are 3 times cheaper to store and 30% more expensive to restore (at the point of this writing, restoring a regular backup costs as much as storing it for 1.5 months, but restoring a cold backup costs as much as storing it for 6.6 months).

When we are ready to implement this feature, we should add here more implementation ideas, and possibly split this issue to several issues. In particular it will be easier for us to implement on-demand backup first - using Scylla's existing snapshot feature - and leave the continuous backup feature as an open issue for later.

In July 2022, DynamoDB published a paper in Usenix ATC. Among other things, the paper suggests how they implementated these of backup features by saving "log files" (what Scylla calls commit logs) to s3, and also periodically creating from them "snapshots" at particular times.