...
Our new digital collection application (scihist_digcol) has slightly different preservation needs than the current Samvera application. However, however our institutional goals remain the same. While we will be making technical changes to preservation methods, our backup and recovery goals are still linked to the following key points in order of importance:
...
Minimal Data Loss: In this case a small subset of data is lost, general site functionality is unaffected.
Minimal Data Inaccessibility: A small subset of data is temporarily inaccessible, general site functionality is unaffected.
Major Data Inaccessibility: Data is not lost permanently, but our ability to access the data is compromised. In this case general site functionality is affected.
Major Data Loss: A large amount of data is lost, and general site functionality is impacted.
Generally speaking, public recovery is the higher initial priority since it impacts more people, contacting staff about outages is easy, and public outages affect the perception of the institute.
Inaccessibility is also different from data loss, though they share certain characteristics. In both cases a solution is to have additional copies of data, but for inaccessibility it is so they can be used as a temporary source of data until the outage is resolved. Handling data inaccessibility requires that the secondary source of data be similarly structured to minimize the time to switch over. With data loss the backups can be in any format as the intent is that a new source of data will be built from the backups.
Our data can be broken into two categories, one is data that is potentially irrecoverable. This includes our original binary files (images, audio, other) and the metadata about them (a SQL database). The other data is restorable but needed for normal site operation but takes significant time to restore, such as the derived download files and viewer files. The second set of data may be worth backing up to shorten recovery times for public users when data is lost.
As an estimate, our cost to hold extra backups for our current scihistcoll staging environment costs less than $2 a month out of a total $70.07 spent on data storage inclusive of data transfer and storage. While a production environment will have slightly higher cost ratios, it should not be massively higher. Thus by spending an additional 2-3% cost on S3, we can mount a full public recovery in an afternoon from a massive failure of our entire infrastructure. While we currently are not backing up our viewer tiles, an examination of our old application shows the cost for production averages around 5 dollars for storage. Adding a second copy of the viewer files should roughly double the cost, with a slight reduction for less use, so will add another 5 dollars to the cost, so for about $7-12 dollars a month we can be widely covered for data inaccessibility or other failures of S3 in a specific region. While it is hard to get specific details, there have been multiple outages or issues in a region whose duration lasted over an hour, and at least one major outage in the last two years lasting around 6 hours. Assuming about 8 hours of problems every two years, we can estimate that a rough cost of $36/hour of outage spent to avoid being down. Shorter outages may not be worth the difficulty of switching over.
In cases of small scale data loss, such as corrupted files or user error, the application will be working fine but a limited set of data will have a problem. In these cases we can locate the problematic data and use a backup copy to restore any damaged original files or use versioning to restore an earlier version of the file. Derived files can either be regenerated or copied from backups as well. This is the most common expected use case at requires only that we keep versions of our original files and backups of files in different locations (another S3 bucket and an on-site copy).
In cases of broad data loss most or all of the data is rendered unavailable. In these cases we will suffer a loss of service until we can recover the data. This can be thought of as two recoveries, : one is to get the digital collection site back as soon as possible for the public and the second is to restore all functionality. Getting the site back for the public is our primary concern, so as noted above for outages we have a few methods to speed up recovery at an additional cost to our backup costs. Both the derivative and original files are backed up to a region on the West Coast in S3 with the same configuration details that our files in US-EAST (the original originals and original derivatives) have. We can recover public access by using these backup files directly while we spend more time working on a full recovery for staff functionality. Due to current setups we would not want staff to add new works, but this allows us to rapidly restore public facing access to our site should the normal data sources be unavailable. A longer process allows us to restore the data back to the original locations while leaving public access up, once the data is restored to its original place full staff functionality will likewise be restored.
...
The first two are the ones that are needed the derivatives are merely backed up because the cost to back them up is low relative to the amount of time saved on a recovery by having a copy ready. If they were lost, the derivatives can be easily recreated based on the first two, but it would take over a day. Thus, we keep a backup anyway because it's cheap and minimizes downtime.
- The postgres database is to be backed up to S3, with a version history of the last 30 versions of the file representing roughly a month of backups roughly.
- The binary files are replicated via S3 replication to a second location in US-WEST rather than US-EAST in case of outages. (During development, as of summer 2019, these binaries are stored at https://s3.console.aws.amazon.com/s3/buckets/scihi-kithe-stage-originals .) When, as part of launching the site, we actually switch over to production, these will also be backed up over to local on-site storage. These The binary files are also versioned and prior versions are held for 30 days before being cleared away to reduce storage costs. This offers a month period to revert a file back if something is damaged.
- The derivative files will also be replicated via S3 replication to a US-WEST location. They can also be regenerated by the application though this takes days to complete for do if all the files are lost. Replication requires versioning, so this is enabled but unlikely to be used.
Minimal public recovery requires the following data:
Postgres database
Original binaries for downloading tiffs
Derivative files
If we have an AWS outage affecting our region, the fastest recovery options are to (if needed) rebuild the servers in another region and edit the local_env.yml file in either ansible or by hand to point to the backup S3 buckets for the original and derivative files. The postgres database will need to be downloaded from S3 and installed onto the new servers if we needed to use new servers. After that point all current public facing aspects will be restored. Since the backup buckets do not sync with the original data buckets, staff should not upload new files though they can edit metadata on existing works. (Once the original S3 buckets have had service restored or their data copied back, set the application to use the original buckets with local_env.yml and users can now add items. The postgres database may need to be copied back to the original server(s) or region if a new server setup was used.
In the case of smaller issues, like single file corruption or deletion, the simplest method for original files is to local locate them on S3 and look at previous versions. We keep 30 days worth of versions so if the error was found within a month you should be able to revert back to an earlier file. For derivatives it is easier to simply regenerate them via the command line.