Backup and Restore

ArangoDB supports three backup methods:

  1. Physical (raw or “cold”) backups
  2. Logical backups
  3. Hot backups

These backup methods save the data which is in the database system. In addition, make sure to backup things like configuration files, startup scripts, Foxx services, access tokens, secrets, certificates etc. and store them in a different location securely.

Performing frequent backups is important and a recommended best practices that can allow you to recover your data in case unexpected problems occur. Hardware failures, system crashes, or users mistakenly deleting data can always happen. Furthermore, while a big effort is put into the development and testing of ArangoDB (in all its deployment modes), ArangoDB, as any other software product, might include bugs or errors and data loss could occur. It is therefore important to regularly backup your data to be able to recover and get up and running again in case of serious problems.

Creating backups of your data before an ArangoDB upgrade is also a best practice.

Making use of a high availability deployment mode of ArangoDB, like Active Failover, Cluster or data-center to data-center replication, does not remove the need of taking frequent backups, which are recommended also when using such deployment modes.

Physical backups

Physical (raw or “cold”) backups can be done when the ArangoDB Server is not running by making a raw copy of the ArangoDB data directory.

Such backups are extremely fast as they only involve file copying.

If ArangoDB is running in Active Failover or Cluster mode, it will be necessary to copy the data directories of all the involved processes (Agents, Coordinators and DB-Servers).

It is extremely important that physical backups are taken only after all the ArangoDB processes have been shut down and the processes are not running anymore. Otherwise files might still be written to, likely resulting in a corrupt and incomplete backup.

It is not always possible to take a physical backup as this method requires a shutdown of the ArangoDB processes. However in some occasions such backups are useful, often in conjunction to the backup coming from another backup method.

Logical Backups

Logical backups can be created and restored with the tools arangodump and arangorestore.

In order to speed up the arangorestore performance in a Cluster environment, the Fast Cluster Restore procedure is recommended.

Hot Backups

Introduced in: v3.5.1

Hot backup and restore associated operations can be performed with the arangobackup client tool and the Hot Backup HTTP API.

Arangobackup and the Hot Backup API are only available in the Enterprise Edition, including ArangoDB Oasis.

Many operations cannot afford downtimes and thus require administrators and operators to create consistent freezes of the data during normal operation. Such use cases imply that near instantaneous hot backups must be obtained in sync across say a cluster’s deployment. For this purpose the hot backup mechanism was created.

The process of creating hot backups is ideally an instantaneous event during normal operations, that consists of a few subsequent steps behind the scenes:

  • Stop all write accesses to the entire installation using a write transaction lock.
  • Create a new local directory under <data-dir>/backups/<timestamp>_<backup-label>.
  • Create hard links to the active database files in <data-dir> in the newly created backup directory.
  • Release the write transaction lock to resume normal operation.
  • Report success of the operation.

The above quite precisely describes the tasks in a single instance installation and could technically finish in under a millisecond. The unknown factor above is of course, when the hot backup process is able to obtain the write transaction lock.

When considering the ArangoDB cluster two more steps need to integrate while others just become slightly more exciting. On the Coordinator tasked with the hot backup the following is done:

  • Using the Agency, make sure that no two hot backups collide.
  • Obtain a dump of the Agency’s Plan key.
  • Stop all write access to the entire cluster installation using a global write transaction lock, this amounts to get each local write transaction lock on each DB-Server, all at the same time.
  • Getting all the locks on the DB-Servers is tried using subsequently growing time periods, and if not all local locks can be acquired during a period, all locks are released again to allow writes to continue. If it is not possible to acquire all local locks in the same period, and this continues for an extended, configurable amount of time, the Coordinator gives up. With the allowInconsistent option set to true, it proceeds instead to create a potentially non-consistent hot backup.
  • On each DB-Server create a new local directory under <data-dir>/backups/<timestamp>_<backup-label>.
  • On each DB-Server create hard links to the active database files in <data-dir> in the newly created backup directory.
  • On each DB-Server store a redundant copy of the above Agency dump.
  • Release the global write transaction lock to resume normal operation.
  • Report success of the operation.

Again under good conditions, a complete hot backup could be obtained from a cluster with many DB-Servers within a very short time in the range of that of the single server installation.

Technical Details

  • The Global Write Transaction Lock

    The global write transaction lock mentioned above is such a determining factor, that it needs a little detailed attention.

    It is obvious that in order to be able to create a consistent snapshot of the ArangoDB world on a specific single server or cluster deployment, one must stop all transactional write operations at the next possible time or else consistency would no longer be given.

    On the other hand it is also obvious, that there is no way for ArangoDB to known, when that time will come. It might be there with the next attempt a nanosecond away, but it could of course not come for the next 2 minutes.

    ArangoDB tries to obtain that lock over and over again. On the single server instances these consecutive tries will not be noticeable. At some point the lock is obtained and the hot backup is created then within a very short amount of time.

    In clusters things are a little more complicated and noticeable. A Coordinator, which is trying to obtain the global write transaction lock must try to get local locks on all DB-Servers simultaneously; potentially succeeding on some and not succeeding on others, leading to apparent dead times in the cluster’s write operations.

    This process can happen multiple times until success is achieved. One has control over the length of the time during which the lock is tried to be obtained each time prolonging the last wait time by 10%.

  • Agency Lock

    Less of a variable, however equally important is to obtain a freeze on the cluster’s structure itself. This is done through the creation of a simple key lock in the cluster’s configuration to stop all ongoing background tasks, which are there to handle fail overs, shard movings, server removals etc. Its role is also to prevent multiple simultaneous hot backup operations. The acquisition of this key is predictably done within a matter of a few seconds.

  • Operation’s Time Scope

    Once the global write transaction lock is obtained, everything goes very quickly. A new backup directory is created, the write ahead lock is flushed and hard links are made on file system level to all persistent files. The duration is not affected by the amount of data in ArangoDB and is near instantaneous.

  • Point in Time Recovery

    One of the great advantages of the method is the consistent snapshot nature. It gives the operator of the database the ability to persist a true and complete time freeze at near zero impact on the ongoing operation. The recovery is easy and restores the entire ArangoDB installation to a desired snapshot.

    Apart from the ability of creating such snapshots it offers a great and easy to use opportunity to experiment with ArangoDB with a means to protect against data loss or corruption.

  • Remote Upload and Download

    We have fully integrated the rclone sync for cloud storage. Rclone is a very versatile inter site sync facility, which opens up a vast field of transport protocols and remote syncing APIs from Amazon’s S3 over Dropbox, WebDAV, all the way to the local file system and network storage.

    One can use the upload and download functionalities to migrate entire cluster installations in this way, copy cluster and single server snapshots all over the world, create an intuitive and easy to use quick access safety backbone of the data operation.

    Rclone is open source and available under the MIT license, is battle tested and has garnered close to 15k stars on GitHub professing to the confidence of lots of users.

Hot Backup Limitations

ArangoDB hot backups impose limitations with respect to storage engine, storage usage, upgrades, deployment scheme, etc. Please review the below list of limitations closely to conclude which operations it might or might not be suited for.

  • Global Scope

    In order to be able to create hot backups instantaneously, they are created on the file system level and thus well below any structural entity related to databases, collections, indexes, users, etc.

    As a consequence, a hot backup is a backup of the entire ArangoDB single server or cluster. In other words, one cannot restore to an older hot backup of a single collection or database. With every restore, one restores the entire deployment including of course the _system database.

    Note that this applies in particular in the case that a certain user might have admin access for the _system database, but explicitly has no access to certain collections. The backup will still extend across all collections!

    It cannot be stressed enough that a restore to an earlier hot backup snapshot will also revert users, graphs, Foxx apps - everything - back to that at the time of the hot backup.

  • Cluster’s Special Limitations

    Creating hot backups can only be done while the internal structure of the cluster remains unaltered. The background of this limitation lies in the distributed nature and the asynchronicity of creation, alteration and dropping of cluster databases, collections and indexes.

    It must be ensured that for the hot backup no such changes are made to the cluster’s inventory, as this could lead to inconsistent hot backups.

  • Restoring from a Different Version

    Hot backups share the same limitations with respect to different versions as ArangoDB itself. This means that a hot backup created with some version a.b.c can without any limitations be restored on any version a.b.d with d not equal to c, that is, the patch level can be changed arbitrarily. With respect to minor versions (second number, b), one can only upgrade and not downgrade. That is, a hot backup created with a version a.b.c can be restored on a version a.d.e for d greater than b but not for d less than b. At this stage, we do not guarantee any compatibility between versions with a different major version number (first number).

  • Identical Topology

    Unlike dumps created with arangodump and restored with arangorestore, hot backups can only be restored to the same type and structure of deployment. This means that one cannot restore a 3-node ArangoDB cluster’s hot backup to any other deployment than another 3-node ArangoDB cluster of the same version.

  • Storage Space

    Without the creation of hot backups, RocksDB keeps compacting the file system level files as the operation continues. Compacted files are subsequently deleted automatically. Every hot backup needs to hold on to the files as they were at the moment of the hot backup creation, thus preventing the deletions and consequently growing the storage space of the ArangoDB data directory. That growth of course depends on the amount of write operations per time.

    This is a crucial factor for sustained operation and might require significantly higher storage reservation for ArangoDB instances involved and a much more fine grained monitoring of storage usage than before.

    Also note that in a cluster each RocksDB instance will be backed up individually and hence the overall storage space will be the sum of all RocksDB instances (i.e., data which is replicated between instances will not be de-duplicated for performance reasons).

  • Global Transaction Lock

    In order to be able to create consistent hot backups, it is mandatory to get a very brief global transaction lock across the entire installation. In single server deployments constant invocation of very long running transactions could prevent that from ever happening during a timeout period. The same holds true for clusters, where this lock must now be obtained on all DB-Servers at the same time.

    Especially in the cluster the result of these successively longer tries to obtain the global transaction lock might become visible in periods of apparent dead time. Locks might be obtained on some machines and and not on others, so that the process has to be retried over and over. Every unsuccessful try would then lead to the release of all partial locks.

    The arangobackup tool provides a --force option since ArangoDB v3.6.0 that can be used to abort ongoing write transactions and thus to more quickly obtain the global transaction lock.

    At this stage, index creation constitutes a write transactions, which means that during index creation one cannot create a hot backup. We intend to lift this limitation in a future version.

  • Services on Single Server

    On a single server the installed Foxx microservices are not backed up and are therefore also not restored. This is because in single server mode the service installation is done locally in the file system and does not track the information in the _apps collection.

    In a cluster, the Coordinators will eventually restore the state of the services from the _apps and _appbundles collections after a backup is restored.

  • Encryption at Rest

    Currently, the hot backup simply takes a snapshot of the database files. If one is using encryption at rest, then the backed up files will be encrypted, with the encryption key that was used in the instance which created the backup.

    Such an encrypted backup can only be restored to an instance using the same encryption key.

  • Replication and Hot Backup

    Hot backups are not automatically replicated between instances. This is true for both the Active Failover setup with 2 (or more) single servers and for the Datacenter to Datacenter Replication between clusters. Simply take hot backups on all instances.

  • Known Issues

    See the list of Known Issues.