| project.description |
|---|
Share data between nodes and perform updates without coordination in an Akka Cluster using Conflict Free Replicated Data Types CRDT. |
You are viewing the documentation for the new actor APIs, to view the Akka Classic documentation, see @ref:Classic Distributed Data.
The Akka dependencies are available from Akka's library repository. To access them there, you need to configure the URL for this repository.
@@repository [sbt,Maven,Gradle] { id="akka-repository" name="Akka library repository" url="https://repo.akka.io/maven" }
To use Akka Cluster Distributed Data, you must add the following dependency in your project:
@@dependency[sbt,Maven,Gradle] { bomGroup=com.typesafe.akka bomArtifact=akka-bom_$scala.binary.version$ bomVersionSymbols=AkkaVersion symbol1=AkkaVersion value1="$akka.version$" group=com.typesafe.akka artifact=akka-cluster-typed_$scala.binary.version$ version=AkkaVersion }
@@project-info{ projectId="akka-cluster-typed" }
Akka Distributed Data is useful when you need to share data between nodes in an Akka Cluster. The data is accessed with an actor providing a key-value store like API. The keys are unique identifiers with type information of the data values. The values are Conflict Free Replicated Data Types (CRDTs).
All data entries are spread to all nodes, or nodes with a certain role, in the cluster via direct replication and gossip based dissemination. You have fine grained control of the consistency level for reads and writes.
The nature of CRDTs makes it possible to perform updates from any node without coordination. Concurrent updates from different nodes will automatically be resolved by the monotonic merge function, which all data types must provide. The state changes always converge. Several useful data types for counters, sets, maps and registers are provided and you can also implement your own custom data types.
It is eventually consistent and geared toward providing high read and write availability (partition tolerance), with low latency. Note that in an eventually consistent system a read may return an out-of-date value.
You can interact with the data through the replicator actor which can be accessed through the @apidoc[typed.*.DistributedData] extension.
The messages for the replicator, such as @apidoc[typed.*Replicator.Update] are defined as
subclasses of @apidoc[typed.*Replicator.Command]
and the actual CRDTs are defined in the akka.cluster.ddata package, for example
@apidoc[akka.cluster.ddata.GCounter]. It requires a @scala[implicit] akka.cluster.ddata.SelfUniqueAddress,
available from:
Scala : @@snip ReplicatorSpec.scala { #selfUniqueAddress }
Java : @@snip ReplicatorTest.java { #selfUniqueAddress }
The replicator can contain multiple entries each containing a replicated data type, we therefore need to create a key identifying the entry and helping us know what type it has, and then use that key for every interaction with the replicator. Each replicated data type contains a factory for defining such a key.
Cluster members with status @ref:WeaklyUp,
will participate in Distributed Data. This means that the data will be replicated to the
WeaklyUp nodes with the background gossip protocol. Note that it
will not participate in any actions where the consistency mode is to read/write from all
nodes or the majority of nodes. The WeaklyUp node is not counted
as part of the cluster. So 3 nodes + 5 WeaklyUp is essentially a
3 node cluster as far as consistent actions are concerned.
This sample uses the replicated data type GCounter to implement a counter that can be written to on any node of the
cluster:
Scala : @@snip ReplicatorSpec.scala { #sample }
Java : @@snip ReplicatorTest.java { #sample }
Although you can interact with the Replicator using the @scala[ActorRef[Replicator.Command]]@java[ActorRef<Replicator.Command>]
from @scala[DistributedData(ctx.system).replicator]@java[DistributedData(ctx.getSystem()).replicator()] it's
often more convenient to use the ReplicatorMessageAdapter as in the above example.
To modify and replicate a data value you send a Replicator.Update message to the local
Replicator.
In the above example, for an incoming Increment command, we send the replicator a Replicator.Update request,
it contains five values:
- the @scala[
Key]@java[KEY] we want to update - the data to use as the empty state if the replicator has not seen the key before
- the @ref:write consistency level we want for the update
- an @scala[
ActorRef[Replicator.UpdateResponse[GCounter]]]@java[ActorRef<Replicator.UpdateResponse<GCounter>>] to respond to when the update is completed - a
modifyfunction that takes a previous state and updates it, in our case by incrementing it with 1
@@@ div { .group-scala }
There is alternative way of constructing the function for the Update message:
Scala : @@snip ReplicatorSpec.scala { #curried-update }
@@@
The current data value for the key of the Update is passed as parameter to the modify
function of the Update. The function is supposed to return the new value of the data, which
will then be replicated according to the given @ref:write consistency level.
The modify function is called by the Replicator actor and must therefore be a pure
function that only uses the data parameter and stable fields from enclosing scope. It must
for example not access the ActorContext or mutable state of an enclosing actor.
Update is intended to only be sent from an actor running in same local ActorSystem
as the Replicator, because the modify function is typically not serializable.
You will always see your own writes. For example if you send two Update messages
changing the value of the same key, the modify function of the second message will
see the change that was performed by the first Update message.
As reply of the Update a Replicator.UpdateSuccess is sent to the replyTo of the
Update if the value was successfully replicated according to the supplied consistency
level within the supplied timeout. Otherwise a Replicator.UpdateFailure subclass is
sent back. Note that a Replicator.UpdateTimeout reply does not mean that the update completely failed
or was rolled back. It may still have been replicated to some nodes, and will eventually
be replicated to all nodes with the gossip protocol.
It is possible to abort the Update when inspecting the state parameter that is passed in to
the modify function by throwing an exception. That happens before the update is performed and
a Replicator.ModifyFailure is sent back as reply.
To retrieve the current value of a data you send Replicator.Get message to the
Replicator.
The example has the GetValue command, which is asking the replicator for current value. Note how the replyTo from the
incoming message can be used when the GetSuccess response from the replicator is received.
@@@ div { .group-scala }
Alternative way of constructing the function for the Get and Delete:
Scala : @@snip ReplicatorSpec.scala { #curried-get }
@@@
For a Get you supply a @ref:read consistency level.
You will always read your own writes. For example if you send a Update message
followed by a Get of the same key the Get will retrieve the change that was
performed by the preceding Update message. However, the order of the reply messages are
not defined, i.e. in the previous example you may receive the GetSuccess before
the UpdateSuccess.
As reply of the Get a Replicator.GetSuccess is sent to the replyTo of the
Get if the value was successfully retrieved according to the supplied consistency
level within the supplied timeout. Otherwise a Replicator.GetFailure is sent.
If the key does not exist the reply will be Replicator.NotFound.
Whenever the distributed counter in the example is updated, we cache the value so that we can answer
requests about the value without the extra interaction with the replicator using the GetCachedValue command.
When we start up the actor we subscribe it to changes for our key, meaning whenever the replicator observes a change
for the counter our actor will receive a @scala[Replicator.Changed[GCounter]]@java[Replicator.Changed<GCounter>]. Since
this is not a message in our protocol, we use a message transformation function to wrap it in the internal InternalSubscribeResponse
message, which is then handled in the regular message handling of the behavior, as shown in the above example.
Subscribers will be notified of changes, if there are any, based on the
configurable akka.cluster.distributed-data.notify-subscribers-interval.
The subscriber is automatically unsubscribed if the subscriber is terminated. A subscriber can
also be de-registered with the replicatorAdapter.unsubscribe(key) function.
In addition to subscribing to individual keys it is possible to subscribe to all keys with a given prefix
by using a * at the end of the key id. For example GCounterKey("counter-*"). Notifications will be
sent for all matching keys, also new keys added later.
A data entry can be deleted by sending a Replicator.Delete message to the local
Replicator. As reply of the Delete a Replicator.DeleteSuccess is sent to
the replyTo of the Delete if the value was successfully deleted according to the supplied
consistency level within the supplied timeout. Otherwise a Replicator.ReplicationDeleteFailure
is sent. Note that ReplicationDeleteFailure does not mean that the delete completely failed or
was rolled back. It may still have been replicated to some nodes, and may eventually be replicated
to all nodes.
A deleted key cannot be reused again, but it is still recommended to delete unused
data entries because that reduces the replication overhead when new nodes join the cluster.
Subsequent Delete, Update and Get requests will be replied with Replicator.DataDeleted.
Subscribers will receive Replicator.Deleted.
The @ref:automatic expiry is an alternative for removing unused data entries.
@@@ warning
As deleted keys continue to be included in the stored data on each node as well as in gossip
messages, a continuous series of updates and deletes of top-level entities will result in
growing memory usage until an ActorSystem runs out of memory. To use Akka Distributed Data
where frequent adds and removes are required, you should use @ref:automatic expiry
or a fixed number of top-level data types that support both updates and removals, for example ORMap or ORSet.
@@@
A data entry can automatically be removed after a period of inactivity, i.e. when there has been no access of
the entry with Get, Update or Delete.
Expiry is enabled for configured keys:
akka.cluster.distributed-data.expire-keys-after-inactivity {
"key-1" = 10 minutes
"cache-*" = 2 minutes
}
Prefix matching is supported by using * at the end of a key.
Expiry can be enabled for all entries by specifying:
akka.cluster.distributed-data.expire-keys-after-inactivity {
"*" = 10 minutes
}
Subscribers will receive Replicator.Expired when an entry has expired.
Expired entries are completely removed and does not leave any tombstones as is the case for @ref:Delete. Expired keys can be reused again. Also @ref:deleted entries can be expired and then completely removed.
The consistency level that is supplied in the @ref:Update and @ref:Get specifies per request how many replicas that must respond successfully to a write and read request.
WriteAll and ReadAll is the strongest consistency level, but also the slowest and with
lowest availability. For example, it is enough that one node is unavailable for a Get request
and you will not receive the value.
For low latency reads you use @scala[ReadLocal]@java[readLocal] with the risk of retrieving stale data, i.e. updates
from other nodes might not be visible yet.
When using @scala[WriteLocal]@java[writeLocal] the Update is only written to the local replica and then disseminated
in the background with the gossip protocol, which can take few seconds to spread to all nodes.
For an update you supply a write consistency level which has the following meaning:
- @scala[
WriteLocal]@java[writeLocal] the value will immediately only be written to the local replica, and later disseminated with gossip WriteTo(n)the value will immediately be written to at leastnreplicas, including the local replicaWriteMajoritythe value will immediately be written to a majority of replicas, i.e. at least N/2 + 1 replicas, where N is the number of nodes in the cluster (or cluster role group)WriteMajorityPlusis likeWriteMajoritybut with the given number ofadditionalnodes added to the majority count. At most all nodes. This gives better tolerance for membership changes between writes and reads. Exiting nodes are excluded usingWriteMajorityPlusbecause those are typically about to be removed and will not be able to respond.WriteAllthe value will immediately be written to all nodes in the cluster (or all nodes in the cluster role group). Exiting nodes are excluded usingWriteAllbecause those are typically about to be removed and will not be able to respond.
When you specify to write to n out of x nodes, the update will first replicate to n nodes.
If there are not enough Acks after a 1/5th of the timeout, the update will be replicated to n other
nodes. If there are less than n nodes left all of the remaining nodes are used. Reachable nodes
are preferred over unreachable nodes.
Note that WriteMajority and WriteMajorityPlus have a minCap parameter that is useful to specify to
achieve better safety for small clusters.
If consistency is a priority, you can ensure that a read always reflects the most recent write by using the following formula:
(nodes_written + nodes_read) > N
where N is the total number of nodes in the cluster, or the number of nodes with the role that is
used for the Replicator.
You supply a consistency level which has the following meaning:
- @scala[
ReadLocal]@java[readLocal] the value will only be read from the local replica ReadFrom(n)the value will be read and merged fromnreplicas, including the local replicaReadMajoritythe value will be read and merged from a majority of replicas, i.e. at least N/2 + 1 replicas, where N is the number of nodes in the cluster (or cluster role group)ReadMajorityPlusis likeReadMajoritybut with the given number ofadditionalnodes added to the majority count. At most all nodes. This gives better tolerance for membership changes between writes and reads. Exiting nodes are excluded usingReadMajorityPlusbecause those are typically about to be removed and will not be able to respond.ReadAllthe value will be read and merged from all nodes in the cluster (or all nodes in the cluster role group). Exiting nodes are excluded usingReadAllbecause those are typically about to be removed and will not be able to respond.
Note that ReadMajority and ReadMajorityPlus have a minCap parameter that is useful to specify to achieve
better safety for small clusters.
When using ReadLocal, you will never receive a GetFailure response, since the local replica is always available to
local readers. WriteLocal however may still reply with UpdateFailure messages if the modify function
throws an exception, or if it fails to persist to @ref:durable storage.
In a 7 node cluster these consistency properties are achieved by writing to 4 nodes and reading from 4 nodes, or writing to 5 nodes and reading from 3 nodes.
By combining WriteMajority and ReadMajority levels a read always reflects the most recent write.
The Replicator writes and reads to a majority of replicas, i.e. N / 2 + 1. For example,
in a 5 node cluster it writes to 3 nodes and reads from 3 nodes. In a 6 node cluster it writes
to 4 nodes and reads from 4 nodes.
You can define a minimum number of nodes for WriteMajority and ReadMajority,
this will minimize the risk of reading stale data. Minimum cap is
provided by minCap property of WriteMajority and ReadMajority and defines the required majority.
If the minCap is higher then N / 2 + 1 the minCap will be used.
For example if the minCap is 5 the WriteMajority and ReadMajority for cluster of 3 nodes will be 3, for
cluster of 6 nodes will be 5 and for cluster of 12 nodes will be 7 ( N / 2 + 1 ).
For small clusters (<7) the risk of membership changes between a WriteMajority and ReadMajority
is rather high and then the nice properties of combining majority write and reads are not
guaranteed. Therefore the ReadMajority and WriteMajority have a minCap parameter that
is useful to specify to achieve better safety for small clusters. It means that if the cluster
size is smaller than the majority size it will use the minCap number of nodes but at most
the total size of the cluster.
In some rare cases, when performing an Update it is needed to first try to fetch latest data from
other nodes. That can be done by first sending a Get with ReadMajority and then continue with
the Update when the GetSuccess, GetFailure or NotFound reply is received. This might be
needed when you need to base a decision on latest information or when removing entries from an ORSet
or ORMap. If an entry is added to an ORSet or ORMap from one node and removed from another
node the entry will only be removed if the added entry is visible on the node where the removal is
performed (hence the name observed-removed set).
@@@ warning
Caveat: Even if you use WriteMajority and ReadMajority there is small risk that you may
read stale data if the cluster membership has changed between the Update and the Get.
For example, in cluster of 5 nodes when you Update and that change is written to 3 nodes:
n1, n2, n3. Then 2 more nodes are added and a Get request is reading from 4 nodes, which
happens to be n4, n5, n6, n7, i.e. the value on n1, n2, n3 is not seen in the response of the
Get request. For additional tolerance of membership changes between writes and reads you can
use WriteMajorityPlus and ReadMajorityPlus.
@@@
For some use cases, for example when limiting the replicator to certain roles, or using different subsets on different roles,
it makes sense to start separate replicators, this needs to be done on all nodes, or
the group of nodes tagged with a specific role. To do this with Distributed Data you will first
have to start a classic Replicator and pass it to the Replicator.behavior method that takes a classic
actor ref. All such Replicators must run on the same path in the classic actor hierarchy.
A standalone ReplicatorMessageAdapter can also be created for a given Replicator instead of creating
one via the DistributedData extension.
Akka contains a set of useful replicated data types and it is fully possible to implement custom replicated data types.
The data types must be convergent (stateful) CRDTs and implement the @scala[ReplicatedData trait]@java[AbstractReplicatedData interface],
i.e. they provide a monotonic merge function and the state changes always converge.
You can use your own custom @scala[ReplicatedData or DeltaReplicatedData]@java[AbstractReplicatedData or AbstractDeltaReplicatedData] types, and several types are provided
by this package, such as:
- Counters:
GCounter,PNCounter - Sets:
GSet,ORSet - Maps:
ORMap,ORMultiMap,LWWMap,PNCounterMap - Registers:
LWWRegister,Flag
GCounter is a "grow only counter". It only supports increments, no decrements.
It works in a similar way as a vector clock. It keeps track of one counter per node and the total
value is the sum of these counters. The merge is implemented by taking the maximum count for
each node.
If you need both increments and decrements you can use the PNCounter (positive/negative counter).
It is tracking the increments (P) separate from the decrements (N). Both P and N are represented
as two internal GCounters. Merge is handled by merging the internal P and N counters.
The value of the counter is the value of the P counter minus the value of the N counter.
Scala : @@snip DistributedDataDocSpec.scala { #pncounter }
Java : @@snip DistributedDataDocTest.java { #pncounter }
GCounter and PNCounter have support for @ref:delta-CRDT and don't need causal
delivery of deltas.
Several related counters can be managed in a map with the PNCounterMap data type.
When the counters are placed in a PNCounterMap as opposed to placing them as separate top level
values they are guaranteed to be replicated together as one unit, which is sometimes necessary for
related data.
Scala : @@snip DistributedDataDocSpec.scala { #pncountermap }
Java : @@snip DistributedDataDocTest.java { #pncountermap }
If you only need to add elements to a set and not remove elements the GSet (grow-only set) is
the data type to use. The elements can be any type of values that can be serialized.
Merge is the union of the two sets.
Scala : @@snip DistributedDataDocSpec.scala { #gset }
Java : @@snip DistributedDataDocTest.java { #gset }
GSet has support for @ref:delta-CRDT and it doesn't require causal delivery of deltas.
If you need add and remove operations you should use the ORSet (observed-remove set).
Elements can be added and removed any number of times. If an element is concurrently added and
removed, the add will win. You cannot remove an element that you have not seen.
The ORSet has a version vector that is incremented when an element is added to the set.
The version for the node that added the element is also tracked for each element in a so
called "birth dot". The version vector and the dots are used by the merge function to
track causality of the operations and resolve concurrent updates.
Scala : @@snip DistributedDataDocSpec.scala { #orset }
Java : @@snip DistributedDataDocTest.java { #orset }
ORSet has support for @ref:delta-CRDT and it requires causal delivery of deltas.
ORMap (observed-remove map) is a map with keys of Any type and the values are ReplicatedData
types themselves. It supports add, update and remove any number of times for a map entry.
If an entry is concurrently added and removed, the add will win. You cannot remove an entry that
you have not seen. This is the same semantics as for the ORSet.
If an entry is concurrently updated to different values the values will be merged, hence the
requirement that the values must be ReplicatedData types.
While the ORMap supports removing and re-adding keys any number of times, the impact that this has on the values
can be non-deterministic. A merge will always attempt to merge two values for the same key, regardless of whether that
key has been removed and re-added in the meantime, an attempt to replace a value with a new one may not have the
intended effect. This means that old values can effectively be resurrected if a node, that has seen both the remove and
the update,gossips with a node that has seen neither. One consequence of this is that changing the value type of the
CRDT, for example, from a GCounter to a GSet, could result in the merge function for the CRDT always failing.
This could be an unrecoverable state for the node, hence, the types of ORMap values must never change for a given key.
It is rather inconvenient to use the ORMap directly since it does not expose specific types
of the values. The ORMap is intended as a low level tool for building more specific maps,
such as the following specialized maps.
ORMultiMap (observed-remove multi-map) is a multi-map implementation that wraps an
ORMap with an ORSet for the map's value.
PNCounterMap (positive negative counter map) is a map of named counters (where the name can be of any type).
It is a specialized ORMap with PNCounter values.
LWWMap (last writer wins map) is a specialized ORMap with LWWRegister (last writer wins register)
values.
ORMap, ORMultiMap, PNCounterMap and LWWMap have support for @ref:delta-CRDT and they require causal
delivery of deltas. Support for deltas here means that the ORSet being underlying key type for all those maps
uses delta propagation to deliver updates. Effectively, the update for map is then a pair, consisting of delta for the ORSet
being the key and full update for the respective value (ORSet, PNCounter or LWWRegister) kept in the map.
Scala : @@snip DistributedDataDocSpec.scala { #ormultimap }
Java : @@snip DistributedDataDocTest.java { #ormultimap }
When a data entry is changed the full state of that entry is replicated to other nodes, i.e.
when you update a map, the whole map is replicated. Therefore, instead of using one ORMap
with 1000 elements it is more efficient to split that up in 10 top level ORMap entries
with 100 elements each. Top level entries are replicated individually, which has the
trade-off that different entries may not be replicated at the same time and you may see
inconsistencies between related entries. Separate top level entries cannot be updated atomically
together.
There is a special version of ORMultiMap, created by using separate constructor
ORMultiMap.emptyWithValueDeltas[A, B], that also propagates the updates to its values (of ORSet type) as deltas.
This means that the ORMultiMap initiated with ORMultiMap.emptyWithValueDeltas propagates its updates as pairs
consisting of delta of the key and delta of the value. It is much more efficient in terms of network bandwidth consumed.
However, this behavior has not been made default for ORMultiMap and if you wish to use it in your code, you
need to replace invocations of ORMultiMap.empty[A, B] (or ORMultiMap()) with ORMultiMap.emptyWithValueDeltas[A, B]
where A and B are types respectively of keys and values in the map.
Please also note, that despite having the same Scala type, ORMultiMap.emptyWithValueDeltas
is not compatible with 'vanilla' ORMultiMap, because of different replication mechanism.
One needs to be extra careful not to mix the two, as they have the same
type, so compiler will not hint the error.
Nonetheless ORMultiMap.emptyWithValueDeltas uses the same ORMultiMapKey type as the
'vanilla' ORMultiMap for referencing.
Note that LWWRegister and therefore LWWMap relies on synchronized clocks and should only be used
when the choice of value is not important for concurrent updates occurring within the clock skew. Read more
in the below section about LWWRegister.
Flag is a data type for a boolean value that is initialized to false and can be switched
to true. Thereafter it cannot be changed. true wins over false in merge.
Scala : @@snip DistributedDataDocSpec.scala { #flag }
Java : @@snip DistributedDataDocTest.java { #flag }
LWWRegister (last writer wins register) can hold any (serializable) value.
Merge of a LWWRegister takes the register with highest timestamp. Note that this
relies on synchronized clocks. LWWRegister should only be used when the choice of
value is not important for concurrent updates occurring within the clock skew.
Merge takes the register updated by the node with lowest address (UniqueAddress is ordered)
if the timestamps are exactly the same.
Scala : @@snip DistributedDataDocSpec.scala { #lwwregister }
Java : @@snip DistributedDataDocTest.java { #lwwregister }
Instead of using timestamps based on System.currentTimeMillis() time it is possible to
use a timestamp value based on something else, for example an increasing version number
from a database record that is used for optimistic concurrency control.
Scala : @@snip DistributedDataDocSpec.scala { #lwwregister-custom-clock }
Java : @@snip DistributedDataDocTest.java { #lwwregister-custom-clock }
For first-write-wins semantics you can use the LWWRegister#reverseClock instead of the
LWWRegister#defaultClock.
The defaultClock is using max value of System.currentTimeMillis() and currentTimestamp + 1.
This means that the timestamp is increased for changes on the same node that occurs within
the same millisecond. It also means that it is safe to use the LWWRegister without
synchronized clocks when there is only one active writer, e.g. a Cluster Singleton. Such a
single writer should then first read current value with ReadMajority (or more) before
changing and writing the value with WriteMajority (or more). When using LWWRegister
with Cluster Singleton it's also recommended to enable:
# Update and Get operations are sent to oldest nodes first.
akka.cluster.distributed-data.prefer-oldest = on
Delta State Replicated Data Types
are supported. A delta-CRDT is a way to reduce the need for sending the full state
for updates. For example adding element 'c' and 'd' to set {'a', 'b'} would
result in sending the delta {'c', 'd'} and merge that with the state on the
receiving side, resulting in set {'a', 'b', 'c', 'd'}.
The protocol for replicating the deltas supports causal consistency if the data type
is marked with RequiresCausalDeliveryOfDeltas. Otherwise it is only eventually
consistent. Without causal consistency it means that if elements 'c' and 'd' are
added in two separate Update operations these deltas may occasionally be propagated
to nodes in a different order to the causal order of the updates. For this example it
can result in that set {'a', 'b', 'd'} can be seen before element 'c' is seen. Eventually
it will be {'a', 'b', 'c', 'd'}.
Note that the full state is occasionally also replicated for delta-CRDTs, for example when new nodes are added to the cluster or when deltas could not be propagated because of network partitions or similar problems.
The the delta propagation can be disabled with configuration property:
akka.cluster.distributed-data.delta-crdt.enabled=off
You can implement your own data types. The only requirement is that it implements
the @scala[merge]@java[mergeData] function of the @scala[ReplicatedData]@java[AbstractReplicatedData] trait.
A nice property of stateful CRDTs is that they typically compose nicely, i.e. you can combine several
smaller data types to build richer data structures. For example, the PNCounter is composed of
two internal GCounter instances to keep track of increments and decrements separately.
Here is s simple implementation of a custom TwoPhaseSet that is using two internal GSet types
to keep track of addition and removals. A TwoPhaseSet is a set where an element may be added and
removed, but never added again thereafter.
Scala : @@snip TwoPhaseSet.scala { #twophaseset }
Java : @@snip TwoPhaseSet.java { #twophaseset }
Data types should be immutable, i.e. "modifying" methods should return a new instance.
Implement the additional methods of @scala[DeltaReplicatedData]@java[AbstractDeltaReplicatedData] if it has support for delta-CRDT replication.
The data types must be serializable with an @ref:Akka Serializer.
It is highly recommended that you implement efficient serialization with Protobuf or similar
for your custom data types. The built in data types are marked with ReplicatedDataSerialization
and serialized with akka.cluster.ddata.protobuf.ReplicatedDataSerializer.
Serialization of the data types are used in remote messages and also for creating message digests (SHA-1) to detect changes. Therefore it is important that the serialization is efficient and produce the same bytes for the same content. For example sets and maps should be sorted deterministically in the serialization.
This is a protobuf representation of the above TwoPhaseSet:
@@snip TwoPhaseSetMessages.proto { #twophaseset }
The serializer for the TwoPhaseSet:
Scala : @@snip TwoPhaseSetSerializer.scala { #serializer }
Java : @@snip TwoPhaseSetSerializer.java { #serializer }
Note that the elements of the sets are sorted so the SHA-1 digests are the same for the same elements.
You register the serializer in configuration:
Scala : @@snip DistributedDataDocSpec.scala { #serializer-config }
Java : @@snip DistributedDataDocSpec.scala { #japi-serializer-config }
Using compression can sometimes be a good idea to reduce the data size. Gzip compression is
provided by the @scala[akka.cluster.ddata.protobuf.SerializationSupport trait]@java[akka.cluster.ddata.protobuf.AbstractSerializationSupport interface]:
Scala : @@snip TwoPhaseSetSerializer.scala { #compression }
Java : @@snip TwoPhaseSetSerializerWithCompression.java { #compression }
The two embedded GSet can be serialized as illustrated above, but in general when composing
new data types from the existing built in types it is better to make use of the existing
serializer for those types. This can be done by declaring those as bytes fields in protobuf:
@@snip TwoPhaseSetMessages.proto { #twophaseset2 }
and use the methods otherMessageToProto and otherMessageFromBinary that are provided
by the SerializationSupport trait to serialize and deserialize the GSet instances. This
works with any type that has a registered Akka serializer. This is how such an serializer would
look like for the TwoPhaseSet:
Scala : @@snip TwoPhaseSetSerializer2.scala { #serializer }
Java : @@snip TwoPhaseSetSerializer2.java { #serializer }
By default the data is only kept in memory. It is redundant since it is replicated to other nodes in the cluster, but if you stop all nodes the data is lost, unless you have saved it elsewhere.
Entries can be configured to be durable, i.e. stored on local disk on each node. The stored data will be loaded next time the replicator is started, i.e. when actor system is restarted. This means data will survive as long as at least one node from the old cluster takes part in a new cluster. The keys of the durable entries are configured with:
akka.cluster.distributed-data.durable.keys = ["a", "b", "durable*"]
Prefix matching is supported by using * at the end of a key.
All entries can be made durable by specifying:
akka.cluster.distributed-data.durable.keys = ["*"]
@scala[LMDB]@java[LMDB] is the default storage implementation. It is
possible to replace that with another implementation by implementing the actor protocol described in
akka.cluster.ddata.DurableStore and defining the akka.cluster.distributed-data.durable.store-actor-class
property for the new implementation.
@@@ note { title="Java 17" }
When using LMDB with Java 17 you have to add JVM flags --add-opens=java.base/sun.nio.ch=ALL-UNNAMED --add-opens=java.base/java.nio=ALL-UNNAMED.
@@@
The location of the files for the data is configured with:
Scala : ```
akka.cluster.distributed-data.durable.lmdb.dir = "ddata"
Java
: ```
# Directory of LMDB file. There are two options:
# 1. A relative or absolute path to a directory that ends with 'ddata'
# the full name of the directory will contain name of the ActorSystem
# and its remote port.
# 2. Otherwise the path is used as is, as a relative or absolute path to
# a directory.
akka.cluster.distributed-data.durable.lmdb.dir = "ddata"
When running in production you may want to configure the directory to a specific path (alt 2), since the default directory contains the remote port of the actor system to make the name unique. If using a dynamically assigned port (0) it will be different each time and the previously stored data will not be loaded.
Making the data durable has a performance cost. By default, each update is flushed
to disk before the UpdateSuccess reply is sent. For better performance, but with the risk of losing
the last writes if the JVM crashes, you can enable write behind mode. Changes are then accumulated during
a time period before it is written to LMDB and flushed to disk. Enabling write behind is especially
efficient when performing many writes to the same key, because it is only the last value for each key
that will be serialized and stored. The risk of losing writes if the JVM crashes is small since the
data is typically replicated to other nodes immediately according to the given WriteConsistency.
akka.cluster.distributed-data.durable.lmdb.write-behind-interval = 200 ms
Note that you should be prepared to receive WriteFailure as reply to an Update of a
durable entry if the data could not be stored for some reason. When enabling write-behind-interval
such errors will only be logged and UpdateSuccess will still be the reply to the Update.
There is one important caveat when it comes pruning of @ref:CRDT Garbage for durable data.
If an old data entry that was never pruned is injected and merged with existing data after
that the pruning markers have been removed the value will not be correct. The time-to-live
of the markers is defined by configuration
akka.cluster.distributed-data.durable.pruning-marker-time-to-live and is in the magnitude of days.
This would be possible if a node with durable data didn't participate in the pruning
(e.g. it was shutdown) and later started after this time. A node with durable data should not
be stopped for longer time than this duration and if it is joining again after this
duration its data should first be manually removed (from the lmdb directory).
There are some limitations that you should be aware of.
CRDTs cannot be used for all types of problems, and eventual consistency does not fit all domains. Sometimes you need strong consistency.
It is not intended for Big Data. The number of top level entries should not exceed 100000. When a new node is added to the cluster all these entries are transferred (gossiped) to the new node. The entries are split up in chunks and all existing nodes collaborate in the gossip, but it will take a while (tens of seconds) to transfer all entries and this means that you cannot have too many top level entries. The current recommended limit is 100000. We will be able to improve this if needed, but the design is still not intended for billions of entries.
All data is held in memory, which is another reason why it is not intended for Big Data.
When a data entry is changed the full state of that entry may be replicated to other nodes if it doesn't support @ref:delta-CRDT. The full state is also replicated for delta-CRDTs, for example when new nodes are added to the cluster or when deltas could not be propagated because of network partitions or similar problems. This means that you cannot have too large data entries, because then the remote message size will be too large.
One thing that can be problematic with CRDTs is that some data types accumulate history (garbage).
For example a GCounter keeps track of one counter per node. If a GCounter has been updated
from one node it will associate the identifier of that node forever. That can become a problem
for long running systems with many cluster nodes being added and removed. To solve this problem
the Replicator performs pruning of data associated with nodes that have been removed from the
cluster. Data types that need pruning have to implement the RemovedNodePruning trait. See the
API documentation of the Replicator for details.
- Strong Eventual Consistency and Conflict-free Replicated Data Types (video) talk by Mark Shapiro
- A comprehensive study of Convergent and Commutative Replicated Data Types paper by Mark Shapiro et. al.
The DistributedData extension can be configured with the following properties:
@@snip reference.conf { #distributed-data }
@java[Distributed Data example project] @scala[Distributed Data example project] is an example project that can be downloaded, and with instructions of how to run.
This project contains several samples illustrating how to use Distributed Data.