Limitations

In General

Transactions in ArangoDB have been designed with particular use cases in mind. They will be mainly useful for short and small data retrieval and/or modification operations.

The implementation is not optimized for very long-running or very voluminous operations, and may not be usable for these cases.

One limitation is that a transaction operation information must fit into main memory. The transaction information consists of record pointers, revision numbers and rollback information. The actual data modification operations of a transaction are written to the write-ahead log and do not need to fit entirely into main memory.

Ongoing transactions will also prevent the write-ahead logs from being fully garbage-collected. Information in the write-ahead log files cannot be written to collection data files or be discarded while transactions are ongoing.

To ensure progress of the write-ahead log garbage collection, transactions should be kept as small as possible, and big transactions should be split into multiple smaller transactions.

Transactions in ArangoDB cannot be nested, i.e. a transaction must not start another transaction. If an attempt is made to call a transaction from inside a running transaction, the server will throw error 1651 (nested transactions detected).

It is also disallowed to execute user transaction on some of ArangoDB’s own system collections. This shouldn’t be a problem for regular usage as system collections will not contain user data and there is no need to access them from within a user transaction.

Some operations are not allowed inside transactions in general:

  • creation and deletion of databases (db._createDatabase(), db._dropDatabase())
  • creation and deletion of collections (db._create(), db._drop(), db.<collection>.rename())
  • creation and deletion of indexes (db.<collection>.ensureIndex(), db.<collection>.dropIndex())

If an attempt is made to carry out any of these operations during a transaction, ArangoDB will abort the transaction with error code 1653 (disallowed operation inside transaction).

Finally, all collections that may be modified during a transaction must be declared beforehand, i.e. using the collections attribute of the object passed to the _executeTransaction function. If any attempt is made to carry out a data modification operation on a collection that was not declared in the collections attribute, the transaction will be aborted and ArangoDB will throw error 1652 unregistered collection used in transaction. It is legal to not declare read-only collections, but this should be avoided if possible to reduce the probability of deadlocks and non-repeatable reads.

In Clusters

Using a single instance of ArangoDB (or a OneShard database in a cluster), multi-document / multi-collection queries are guaranteed to be fully ACID in the traditional sense. For more details see Operation Atomicity and Transactional Isolation. This is more than many other NoSQL database systems support. In cluster mode, single-document operations are also fully ACID.

Multi-document / multi-collection queries and transactions offer different guarantees. Understanding these differences is important when designing applications that need to be resilient against outages of individual servers.

Cluster transactions share the underlying characteristics of the storage engine that is used for the cluster deployment. A transaction started on a Coordinator translates to one transaction per involved DB-Server. The guarantees and characteristics of the given storage-engine apply additionally to the cluster specific information below. Please refer to Locking and Isolation for more details on the storage-engines.

Atomicity

A transaction on one DB-Server is either committed completely or not at all.

ArangoDB transactions do currently not require any form of global consensus. This makes them relatively fast, but also vulnerable to unexpected server outages.

Should a transaction involve Leader Shards on multiple DB-Servers, the atomicity of the distributed transaction during the commit operation can not be guaranteed. Should one of the involved DB-Servers fail during the commit the transaction is not rolled-back globally, sub-transactions may have been committed on some DB-Servers, but not on others. Should this case occur the client application will see an error.

An improved failure handling issue might be introduced in future versions.

Consistency

We provide consistency even in the cluster, a transaction will never leave the data in an incorrect or corrupt state.

In ArangoDB there is always exactly one DB-Server responsible for a given shard. In both Storage-Engines the locking procedures ensure that dependent transactions (in the sense that the transactions modify the same documents or unique index entries) are ordered sequentially. Therefore we can provide Causal-Consistency for your transactions.

From the applications point-of-view this also means that a given transaction can always read it’s own writes. Other concurrent operations will not change the database state seen by a transaction.

Isolation

The ArangoDB Cluster provides Local Snapshot Isolation. This means that all operations and queries in the transactions will see the same version, or snapshot, of the data on a given DB-Server. This snapshot is based on the state of the data at the moment in time when the transaction begins on that DB-Server.

Durability

It is guaranteed that successfully committed transactions are persistent. Using replication and / or waitForSync increases the durability (Just as with the single-server).

RocksDB storage engine

The following restrictions and limitations do not apply to JavaScript transactions, since their intended use case is for smaller transactions with full transactional guarantees. So the following only applies to AQL queries and transactions created through the document API (i.e. batch operations).

Data of ongoing transactions is stored in RAM. Transactions that get too big (in terms of number of operations involved or the total size of data created or modified by the transaction) will be committed automatically. Effectively this means that big user transactions are split into multiple smaller RocksDB transactions that are committed individually. The entire user transaction will not necessarily have ACID properties in this case.

The following global options can be used to control the RAM usage and automatic intermediate commits for the RocksDB engine:

--rocksdb.max-transaction-size

Transaction size limit (in bytes). Transactions store all keys and values in RAM, so large transactions run the risk of causing out-of-memory situations. This setting allows you to ensure that does not happen by limiting the size of any individual transaction. Transactions whose operations would consume more RAM than this threshold value will abort automatically with error 32 (“resource limit exceeded”).

--rocksdb.intermediate-commit-size

If the size of all operations in a transaction reaches this threshold, the transaction is committed automatically and a new transaction is started. The value is specified in bytes.

--rocksdb.intermediate-commit-count

If the number of operations in a transaction reaches this value, the transaction is committed automatically and a new transaction is started.

The above values can also be adjusted per query, by setting the following attributes in the call to db._query():

  • maxTransactionSize: transaction size limit in bytes
  • intermediateCommitSize: maximum total size of operations after which an intermediate commit is performed automatically
  • intermediateCommitCount: maximum number of operations after which an intermediate commit is performed automatically

Limits for Stream transactions

A maximum lifetime and transaction size for Stream Transactions is enforced on the Coordinator to ensure that abandoned transactions cannot block the cluster from operating properly:

  • Maximum idle timeout of up to 120 seconds between operations
  • Maximum transaction size of 128 MB per DB-Server

These limits are also enforced for Stream Transactions on single servers.

The default maximum idle timeout is 60 seconds between operations in a single Stream Transaction. The maximum value can be bumped up to at most 120 seconds by setting the startup option --transaction.streaming-idle-timeout. Posting an operation into a non-expired Stream Transaction will reset the transaction’s timeout to the configured idle timeout.

Enforcing the limit is useful to free up resources used by abandoned transactions, for example from transactions that are abandoned by client applications due to programming errors or that were left over because client connections were interrupted.