To persist an actor's state, only changes to that actor's state are written to a journal, not current state directly. These changes are appended as immutable facts to a journal, nothing is ever mutated, which allows for very high transaction rates and efficient replication. Actor state can be recovered by replaying stored changes and projecting them again. This not only allows state recovery after an actor has been restarted by a supervisor but also after JVM or node crashes, for example. State changes are defined in terms of messages an actor receives (or generates).
Persistence of messages also forms the basis for supporting at-least-once message delivery semantics. This requires retries to counter transport losses, which means keeping state at the sending end and having an acknowledgement mechanism at the receiving end (see Message Delivery Guarantees in Akka). Akka Persistence supports that for point-to-point communications. Reliable point-to-point communications are an important part of highly scalable applications (see also Pet Helland's position paper Life beyond Distributed Transactions).
The following gives a high-level overview of the current features in Akka Persistence. Links to more detailed documentation are included.
ProcessorsProcessors are persistent actors. They internally communicate with a journal to persist messages they receive or generate. They may also request message replay from a journal to recover internal state in failure cases. Processors may either persist messages
- before an actor's behavior is executed (command sourcing) or
- during an actor's behavior is executed (event sourcing)
Event sourced processors do not persist commands. Instead they allow application code to derive events from a command and atomically persist these events. After persistence, they are applied to current state. During recovery, events are replayed and only the state-changing behavior of an event sourced processor is executed again. Other side effects that have been executed during normal operation are not performed again.
Processors automatically recover themselves. Akka Persistence guarantees that new messages sent to a processor never interleave with replayed messages. New messages are internally buffered until recovery completes, hence, an application may send messages to a processor immediately after creating it.
SnapshotsThe recovery time of a processor increases with the number of messages that have been written by that processor. To reduce recovery time, applications may take snapshots of processor state which can be used as starting points for message replay. Usage of snapshots is optional and only needed for optimization.
ChannelsChannels are actors that provide at-least-once message delivery semantics between a sending processor and a receiver that acknowledges the receipt of messages on application level. They also ensure that successfully acknowledged messages are not delivered again to receivers during processor recovery (i.e. replay of messages). Applications that want to have reliable message delivery without application-defined sending processors should use persistent channels. A persistent channel is like a normal channel that additionally persists messages before sending them to a receiver.
JournalsJournals (and snapshot stores) are pluggable in Akka Persistence. The default journal plugin is backed by LevelDB which writes messages to the local filesystem. A replicated journal is planned but not yet part of the distribution. Replicated journals allow stateful actors to be migrated in a cluster, for example. For testing purposes, a remotely shared LevelDB journal can be used instead of a replicated journal to experiment with stateful actor migration. Application code doesn't need to change when switching to a replicated journal later.
Update: A replicated journal backed by Apache Cassandra is available here.