Distributed Go actor framework to build reactive and distributed system in golang using protocol buffers as actor messages.
GoAkt is highly scalable and available when running in cluster mode. It comes with the necessary features require to build a distributed actor-based system without sacrificing performance and reliability. With GoAkt, you can instantly create a fast, scalable, distributed system across a cluster of computers.
If you are not familiar with the actor model, the blog post from Brian Storti here is an excellent and short introduction to the actor model. Also, check reference section at the end of the post for more material regarding actor model.
- Design Principles
- Use Cases
- Installation
- Versioning
- Examples
- Features
- API
- Client
- Clustering
- Contribution
- Benchmark
This framework has been designed:
- to be very simple - it caters for the core component of an actor framework as stated by the father of the actor framework here.
- to be very easy to use.
- to have a clear and defined contract for messages - no need to implement/hide any sort of serialization.
- to make use existing battle-tested libraries in the go ecosystem - no need to reinvent solved problems.
- to be very fast.
- to expose interfaces for custom integrations rather than making it convoluted with unnecessary features.
- Event-Driven programming
- Event Sourcing and CQRS - eGo
- Highly Available, Fault-Tolerant Distributed Systems
go get github.com/tochemey/goakt/v2
The version system adopted in GoAkt deviates a bit from the standard semantic versioning system. The version format is as follows:
- The
MAJOR
part of the version will stay atv2
for the meantime. - The
MINOR
part of the version will cater for any new features, breaking changes with a note on the breaking changes. - The
PATCH
part of the version will cater for dependencies upgrades, bug fixes, security patches and co.
The versioning will remain like v2.x.x
until further notice.
Kindly check out the examples' repository.
The fundamental building blocks of GoAkt are actors.
- They are independent, isolated unit of computation with their own state.
- They can be long-lived actors or be passivated after some period of time that is configured during their creation. Use this feature with care when dealing with persistent actors (actors that require their state to be persisted).
- They are automatically thread-safe without having to use locks or any other shared-memory synchronization mechanisms.
- They can be stateful and stateless depending upon the system to build.
- Every actor in GoAkt:
- has a process id
PID
. Via the process id any allowable action can be executed by the actor. - has a lifecycle via the following methods:
PreStart
,PostStop
.PreStart
hook is used to initialise actor state. It is like the actor constructor.PostStop
hook is used to clean up resources used by the Actor. It is like the actor destructor. It means it can live and die like any other process.
- handles and responds to messages via the method
Receive
. While handling messages it can: - create other (child) actors via their process id
PID
SpawnChild
method - send messages to other actors locally or remotely via their process
id
PID
Ask
,RemoteAsk
(request/response fashion) andTell
,RemoteTell
(fire-and-forget fashion) methods - stop (child) actors via their process id
PID
- watch/unwatch (child) actors via their process id
PID
Watch
andUnWatch
methods - supervise the failure behavior of (child) actors.
- remotely lookup for an actor on another node via their process id
PID
RemoteLookup
. This allows it to send messages remotely viaRemoteAsk
orRemoteTell
methods - stash/unstash messages. See Stashing
- can adopt various form using the Behavior feature
- can be restarted (respawned)
- can be gracefully stopped (killed). Every message in the mailbox prior to stoppage will be processed within a configurable time period.
- has a process id
Actors can be passivated when they are idle after some period of time. Passivated actors are removed from the actor system to free-up resources.
When cluster mode is enabled, passivated actors are removed from the entire cluster. To bring back such actors to live, one needs to
Spawn
them again. By default, all actors are passivated and the passivation time is two minutes
.
- To enable passivation use the actor system option
WithExpireActorAfter(duration time.Duration)
when creating the actor system. See actor system options. - To disable passivation use the actor system option
WithPassivationDisabled
when creating the actor system. See actor system options.
In GoAkt, supervision allows to define the various strategies to apply when a given actor is faulty.
The supervisory strategies to adopt are to be set during the creation of the actor. One can set as many as he/she wants supervisor strategies based upon various error types using the SpawnOption
method WithSupervisorStrategies
.
To create a supervisor strategy one needs to call the NewSupervisorStrategy
function and pass the error type and the corresponding directive.
GoAkt comes bundled with the following default supervisor strategies that can be overriden when creating an actor:
NewSupervisorStrategy(PanicError{}, NewStopDirective())
: this will stop the faulty actor in case of panic.NewSupervisorStrategy(&runtime.PanicNilError{}, NewStopDirective())
: this will stop the faulty actor for nil panic error.
Note: GoAkt will suspend a faulty actor when there is no supervisor strategy set in place for the corresponding error type. Once can check the state of the actor using the IsSuspended
method on the PID
.
A suspended actor can be restarted or shutdown, however it cannot handle any messages sent to it. There are buit-in error types one can use:
PanicError
: to wrap up an error instead of using the go standard panic.InternalError
: it is a wrapper around an error that can be categorized as internal.SpawnError
: recommended when the creation of an actor failed.
In GoAkt each child actor is treated separately. There is no concept of one-for-one and one-for-all strategies. The following directives are supported:
Restart
: to restart the child actor. One can control how the restart is done using the following options: -maxNumRetries
: defines the maximum of restart attempts -timeout
: how to attempt restarting the faulty actor.Stop
: to stop the child actor which is the default one as long as its descendants.Resume
: ignores the failure and process the next message, instead.
With the Restart
directive, only the direct alive children of the given actor will be shudown and respawned with their initial state.
There are only two scenarios where an actor can supervise another actor:
- It watches the given actor via the
Watch
method. With this method the parent actor can also listen to theTerminated
message to decide what happens next to the child actor. - The actor to be supervised is a child of the given actor.
Without an actor system, it is not possible to create actors in GoAkt. Only a single actor system
is recommended to be created per application when using GoAkt. At the moment the single instance is not enforced in GoAkt, this simple implementation is left to the discretion of the developer. To
create an actor system one just need to use
the NewActorSystem
method with the various Options. GoAkt
ActorSystem has the following characteristics:
- Actors lifecycle management (Spawn, Kill, ReSpawn)
- Concurrency and Parallelism - Multiple actors can be managed and execute their tasks independently and concurrently. This helps utilize multicore processors efficiently.
- Location Transparency - The physical location of actors is abstracted when cluster mode is enabled. Remote actors can be accessed via their address once remoting is enabled.
- Fault Tolerance and Supervision - Set during the creation of the actor system.
- Actor Addressing - Every actor in the ActorSystem has an address.
Actors in GoAkt have the power to switch their behaviors at any point in time. When you change the actor behavior, the new behavior will take effect for all subsequent messages until the behavior is changed again. The current message will continue processing with the existing behavior. You can use Stashing to reprocess the current message with the new behavior.
To change the behavior, call the following methods on the ReceiveContext interface when handling a message:
Become
- switches the current behavior of the actor to a new behavior.UnBecome
- resets the actor behavior to the default one which is the Actor.Receive method.BecomeStacked
- sets a new behavior to the actor to the top of the behavior stack, while maintaining the previous ones.UnBecomeStacked()
- sets the actor behavior to the previous behavior beforeBecomeStacked()
was called. This only works withBecomeStacked()
.
Routers help send the same type of message to a set of actors to be processed in parallel depending upon the type of the router used. Routers should be used with caution because they can hinder performance. When the router receives a message to broadcast, every routee is checked whether alive or not. When a routee is not alive the router removes it from its set of routees. When the last routee stops the router itself stops.
GoAkt comes shipped with the following routing strategies:
Fan-Out
: This strategy broadcasts the given message to all its available routees in parallel.Random
: This strategy randomly picks a routee in its set of routees and send the message to it.Round-Robin
: This strategy sends messages to its routee in a round-robin way. For n messages sent through the router, each actor is forwarded one message.
A router a just like any other actor that can be spawned. To spawn router just call the ActorSystem SpawnRouter
method.
Router as well as their routees are not passivated.
Once can implement a custom mailbox. See Mailbox. GoAkt comes with the following mailboxes built-in:
UnboundedMailbox
: this is the default mailbox. It is implemented using the lock-free Multi-Producer-Single-Consumer Queue.BoundedMailbox
: this is a thread-safe mailbox implemented using the Ring-Buffer Queue. When the mailbox is full any new message is sent to the deadletter queue. Setting a reasonable capacity for the queue can enhance throughput.UnboundedPriorityMailbox
: this is thread-safe mailbox using the standard library container/heap. At the moment the performance of is this mailbox is not comparable to the two other built-in mailboxes.
To receive some system events and act on them for some particular business cases, you just need to call the actor system Subscribe
.
Make sure to Unsubscribe
whenever the subscription is no longer needed to free allocated resources.
The subscription methods can be found on the ActorSystem
interface.
ActorStarted
: emitted when an actor has startedActorStopped
: emitted when an actor has stoppedActorPassivated
: emitted when an actor is passivatedActorChildCreated
: emitted when a child actor is createdActorRestarted
: emitted when an actor has restartedActorSuspended
: emitted when an actor has been suspendedNodeJoined
: cluster event emitted when a node joins the cluster. This only happens when cluster mode is enabledNodeLeft
: cluster event emitted when a node leaves the cluster. This only happens when cluster mode is enabledDeadletter
: emitted when a message cannot be delivered or that were not handled by a given actor. Dead letters are automatically emitted when a message cannot be delivered to actors' mailbox or when an Ask times out. Also, one can emit dead letters from the receiving actor by using thectx.Unhandled()
method. This is useful instead of panicking when the receiving actor does not know how to handle a particular message. Dead letters are not propagated over the network, there are tied to the local actor system.
Communication between actors is achieved exclusively through message passing. In GoAkt Google Protocol Buffers is used to define messages. The choice of protobuf is due to easy serialization over wire and strong schema definition. As stated previously the following messaging patterns are supported:
Tell/RemoteTell
- send a message to an actor and forget it.Tell
is used for local messaging.Ask/RemoteAsk
- send a message to an actor and expect a reply within a time period.Ask
is used for local messaging.SendAsync
- behave the same way asTell
. This call is location transparent which means that the system will locate the given actor whether locally or remotely to send the message. This is possible when cluster mode is enabled.SendSync
- behave the same way asAsks
except the location of the provided actor is transparent. This is possible when cluster mode is enabled.Forward
- pass a message from one actor to the actor by preserving the initial sender of the message. At the moment you can only forward messages from theReceiveContext
when handling a message within an actor and this to a local actor.ForwardTo
- behave the same asForward
but when cluster mode is enabled.BatchTell
- send a bulk of messages to actor in a fire-forget manner. Messages are processed one after the other in the other they have been sent.BatchAsk
- send a bulk of messages to an actor and expect responses for each message sent within a time period. Messages are processed one after the other in the other they were sent. This help return the response of each message in the same order that message was sent. This method hinders performance drastically when the number of messages to sent is high. Kindly use this method with caution.PipeTo
- send the successful result of a future(long-running task) to self or a given actor. This can be achieved from thePID
as well as from the ReceiveContext
You can schedule sending messages to actor that will be acted upon in the future. To achieve that you can use the following methods on the Actor System:
ScheduleOnce
- will send the given message to a local actor once after a given intervalSchedule
- will send the given message to a local actor at a given intervalRemoteSchedule
- will send the given message to a remote actor at a given interval. This requires remoting to be enabled on the actor system.RemoteScheduleOnce
- will send the given message to a remote actor once after a given interval. This requires remoting to be enabled on the actor system.ScheduleWithCron
- will send the given message to a local actor using a cron expression.RemoteScheduleWithCron
- will send the given message to a remote actor using a cron expression. This requires remoting to be enabled on the actor system.
Field | Required | Allowed Values | Allowed Special Characters |
---|---|---|---|
Seconds | yes | 0-59 | , - * / |
Minutes | yes | 0-59 | , - * / |
Hours | yes | 0-23 | , - * / |
Day of month | yes | 1-31 | , - * ? / |
Month | yes | 1-12 or JAN-DEC | , - * / |
Day of week | yes | 1-7 or SUN-SAT | , - * ? / |
Year | no | empty, 1970- | , - * / |
When running the actor system in a cluster only one instance of a given scheduled message will be running across the entire cluster.
Stashing is a mechanism you can enable in your actors, so they can temporarily stash away messages they cannot or should not handle at the moment. Another way to see it is that stashing allows you to keep processing messages you can handle while saving for later messages you can't. Stashing are handled by GoAkt out of the actor instance just like the mailbox, so if the actor dies while processing a message, all messages in the stash are processed. This feature is usually used together with Become/UnBecome, as they fit together very well, but this is not a requirement.
It’s recommended to avoid stashing too many messages to avoid too much memory usage. If you try to stash more messages than the capacity the actor will panic. To use the stashing feature, call the following methods on the ReceiveContext when handling a message:
Stash()
- adds the current message to the stash buffer.Unstash()
- unstashes the oldest message in the stash and prepends to the stash buffer.UnstashAll()
- unstashes all messages from the stash buffer and prepends in the mailbox. Messages will be processed in the same order they arrived. The stash buffer will be empty after processing all messages, unless an exception is thrown or messages are stashed while unstashing.
GoAkt provides a mechanism to execute user-defined tasks during the shutdown of the actor system. One can use the option WithCoordinatedShutdown
when creating the actor system. The tasks are executed in the order they have been set. Any failure will halt the shutdown process.
Remoting allows remote actors to communicate. The underlying technology is gRPC. To enable remoting just use the WithRemoting
option when
creating the actor system. See actor system options. These are the following remoting features available:
RemoteTell
: to send a fire-and-forget message to an actor remotelyRemoteAsk
: to send a request/response type of message to a remote actorRemoteBatchTell
: to send a fire-and-forget bulk of messages to a remote actorRemoteBatchAsk
: to send a bulk messages to a remote actor with repliesRemoteLookup
: to lookup for an actor on a remote hostRemoteReSpawn
: to restarts an actor on a remote machineRemoteStop
: to stop an actor on a remote machineRemoteSpawn
: to start an actor on a remote machine. The given actor implementation must be registered using theRegister
method of the actor system on the remote machine for this call to succeed.RemoteForward
: to pass a message from one actor to the actor by preserving the initial sender of the message.
These methods can be found as well as on the PID which is the actor reference when an actor is created.
This offers simple scalability, partitioning (sharding), and re-balancing out-of-the-box. GoAkt nodes are automatically discovered. See Clustering. Beware that at the moment, within the cluster the existence of an actor is unique.
Observability is key in distributed system. It helps to understand and track the performance of a system. GoAkt offers out of the box features that can help track, monitor and measure the performance of a GoAkt based system.
The following methods have been implemented to help push some metrics to any observability tool:
- Total Number of children at a given point in time PID
- Number of messages stashed at a given point in time PID
- Number of Restarts at a given point in time PID
- Latest message received processing duration in milliseconds PID
- Total Number of Actors at a given point in time ActorSystem
A simple logging interface to allow custom logger to be implemented instead of using the default logger.
GoAkt does not support at the moment any form of data encryption or TLS to prevent any form of mitm attack. This feature may come in the future. At the moment, I will recommend a GoAkt-based application should be deployed behind a vpc or using a service mesh like Linkerd or Istio which offers great mTLS support when it comes to service communucation.
GoAkt comes packaged with a testkit that can help test that actors receive expected messages within unit tests.
The teskit in GoAkt uses underneath the https://github.com/stretchr/testify
package.
To test that an actor receive and respond to messages one will have to:
- Create an instance of the testkit:
testkit := New(ctx, t)
wherectx
is a go context andt
the instance of*testing.T
. This can be done in setup before the run of each test. - Create the instance of the actor under test. Example:
pinger := testkit.Spawn(ctx, "pinger", &pinger{})
- Create an instance of test probe:
probe := testkit.NewProbe(ctx)
wherectx
is a go context. One can set some options - Use the probe to send a message to the actor under test. Example:
probe.Send(pinger, new(testpb.Ping))
for a Tell assertion andprobe.SendSync(pinger, new(testpb.Ping), time.Second)
for an Ask assertion. - Assert that the actor under test has received the message and responded as expected using the probe methods:
ExpectMessage(message proto.Message)
: asserts that the message received from the test actor is the expected oneExpectMessageWithin(duration time.Duration, message proto.Message)
: asserts that the message received from the test actor is the expected one within a time durationExpectNoMessage()
: asserts that no message is expectedExpectAnyMessage() proto.Message
: asserts that any message is expectedExpectAnyMessageWithin(duration time.Duration) proto.Message
: asserts that any message within a time durationExpectMessageOfType(messageType protoreflect.MessageType)
: asserts the expectation of a given message typeExpectMessageOfTypeWithin(duration time.Duration, messageType protoreflect.MessageType)
: asserts the expectation of a given message type within a time duration
- Make sure to shut down the testkit and the probe. Example:
probe.Stop()
,testkit.Shutdown(ctx)
wherectx
is a go context. These two calls can be in a tear down after all tests run.
To help implement unit tests in GoAkt-based applications. See Testkit
The API interface helps interact with a GoAkt actor system as kind of client. The following features are available:
Tell
: to send a message to an actor in a fire-and-forget manner.Ask
: to send a message to an actor and expect a response within a given timeout.BatchAsk
: to send a batch of requests to an actore remotely and expect responses back for each request.BatchTell
: to send a batch of fire-and-forget messages to an actor remotely.RemoteTell
: to send a fire-and-forget message to an actor remotely using the Remoting API.RemoteAsk
: to send a request/response type of message to a remote actor using the Remoting API.RemoteBatchTell
: to send a fire-and-forget bulk of messages to a remote actor using the Remoting API.RemoteBatchAsk
: to send a bulk messages to a remote actor with replies using the Remoting API.RemoteLookup
: to lookup for an actor on a remote host using the Remoting API.RemoteReSpawn
: to restarts an actor on a remote machine using the Remoting API.RemoteStop
: to stop an actor on a remote machine using the Remoting API.RemoteSpawn
: to start an actor on a remote machine using the Remoting API. The given actor implementation must be registered using theRegister
method of the actor system on the remote machine for this call to succeed.
The GoAkt client facilitates interaction with a specified GoAkt cluster, contingent upon the activation of cluster mode. The client operates without knowledge of the specific node within the cluster that will process the request. This feature is particularly beneficial when interfacing with a GoAkt cluster from an external system. GoAkt client is equipped with a mini load-balancer that helps route requests to the appropriate node.
- Round Robin - a given node is chosen using the round-robin strategy
- Random - a given node is chosen randomly
- Least Load - the node with the least number of actors is chosen
Kinds
- to list all the actor kinds in the clusterSpawn
- to spawn an actor in the clusterSpawnWithBalancer
- to spawn an actor in the cluster with a given balancer strategyStop
- to kill/stop an actor in the clusterAsk
- to send a message to a given actor in the cluster and expect a responseTell
- to send a fire-forget message to a given actor in the clusterWhereis
- to locate and get the address of a given actorReSpawn
- to restart a given actor
The cluster engine depends upon the discovery mechanism to find other nodes in the cluster.
Under the hood, it leverages Olric
to scale out and guarantee performant, reliable persistence, simple scalability, partitioning (sharding), and
re-balancing out-of-the-box. It requires remoting to be enabled. One can implement a custom hasher for the partitioning using
the Hasher interface and the Actor System
option to set it. The default hasher uses the XXH3 algorithm
.
At the moment the following providers are implemented:
- kubernetes api integration is fully functional
- mDNS and DNS-SD
- NATS integration is fully functional
- DNS is fully functional
- Static is fully functional and for demo purpose
Note: One can add additional discovery providers using the following discovery provider.
The following outlines the cluster mode operations which can help have a healthy GoAkt cluster:
- One can start a single node cluster or a multiple nodes cluster.
- One can add more nodes to the cluster which will automatically discover the cluster.
- One can remove nodes. However, to avoid losing data, one need to scale down the cluster to the minimum number of nodes which started the cluster.
When a node leaves the cluster, as long as the cluster quorum is stable, its actors are redeployed on the remaining nodes of the cluster.
The redeployed actors are created with their initial state. Every field of the Actor set using the PreStart
will have their value set
as expected. On the contrary every field of the Actor will be set to their default go type value because actors are created using reflection.
To get the kubernetes discovery working as expected, the following pod labels need to be set:
app.kubernetes.io/part-of
: set this label with the actor system nameapp.kubernetes.io/component
: set this label with the application nameapp.kubernetes.io/name
: set this label with the application name
package main
import "github.com/tochemey/goakt/v2/discovery/kubernetes"
const (
namespace = "default"
applicationName = "accounts"
actorSystemName = "AccountsSystem"
discoveryPortName = "discovery-port"
peersPortName = "peers-port"
remotingPortName = "remoting-port"
)
// define the discovery config
config := kubernetes.Config{
ApplicationName: applicationName,
ActorSystemName: actorSystemName,
Namespace: namespace,
DiscoveryPortName: gossipPortName,
RemotingPortName: remotingPortName,
PeersPortName: peersPortName,
}
// instantiate the k8 discovery provider
disco := kubernetes.NewDiscovery(&config)
// pass the service discovery when enabling cluster mode in the actor system
You’ll also have to grant the Service Account that your pods run under access to list pods. The following configuration can be used as a starting point. It creates a Role, pod-reader, which grants access to query pod information. It then binds the default Service Account to the Role by creating a RoleBinding. Adjust as necessary:
kind: Role
apiVersion: rbac.authorization.k8s.io/v1
metadata:
name: pod-reader
rules:
- apiGroups: [""] # "" indicates the core API group
resources: ["pods"]
verbs: ["get", "watch", "list"]
---
kind: RoleBinding
apiVersion: rbac.authorization.k8s.io/v1
metadata:
name: read-pods
subjects:
# Uses the default service account. Consider creating a new one.
- kind: ServiceAccount
name: default
roleRef:
kind: Role
name: pod-reader
apiGroup: rbac.authorization.k8s.io
A working example can be found here
ServiceName
: the service nameService
: the service typeDomain
: the mDNS discovery domainPort
: the mDNS discovery portIPv6
: it states whether to lookup for IPv6 addresses.
To use the NATS discovery provider one needs to provide the following:
NatsServer
: the NATS Server addressNatsSubject
: the NATS subject to useActorSystemName
: the actor system nameApplicationName
: the application nameTimeout
: the nodes discovery timeoutMaxJoinAttempts
: the maximum number of attempts to connect an existing NATs server. Defaults to5
ReconnectWait
: the time to backoff after attempting a reconnect to a server that we were already connected to previously. Default to2 seconds
Host
: the given node host addressDiscoveryPort
: the discovery port of the given node
package main
import "github.com/tochemey/goakt/v2/discovery/nats"
const (
natsServerAddr = "nats://127.0.0.1:4248"
natsSubject = "goakt-gossip"
applicationName = "accounts"
actorSystemName = "AccountsSystem"
)
// define the discovery options
config := nats.Config{
ApplicationName: applicationName,
ActorSystemName: actorSystemName,
NatsServer: natsServerAddr,
NatsSubject: natsSubject,
Host: "127.0.0.1",
DiscoveryPort: 20380,
}
// instantiate the NATS discovery provider by passing the config and the hostNode
disco := nats.NewDiscovery(&config)
// pass the service discovery when enabling cluster mode in the actor system
This provider performs nodes discovery based upon the domain name provided. This is very useful when doing local development using docker.
To use the DNS discovery provider one needs to provide the following:
DomainName
: the domain nameIPv6
: it states whether to lookup for IPv6 addresses.
package main
import "github.com/tochemey/goakt/v2/discovery/dnssd"
const domainName = "accounts"
// define the discovery options
config := dnssd.Config{
dnssd.DomainName: domainName,
dnssd.IPv6: false,
}
// instantiate the dnssd discovery provider
disco := dnssd.NewDiscovery(&config)
// pass the service discovery when enabling cluster mode in the actor system
A working example can be found here
This provider performs nodes discovery based upon the list of static hosts addresses.
The address of each host is the form of host:port
where port
is the gossip protocol port.
package main
import "github.com/tochemey/goakt/v2/discovery/static"
// define the discovery configuration
config := static.Config{
Hosts: []string{
"node1:3322",
"node2:3322",
"node3:3322",
},
}
// instantiate the dnssd discovery provider
disco := static.NewDiscovery(&config)
// pass the service discovery when enabling cluster mode in the actor system
A working example can be found here
Contributions are welcome! The project adheres to Semantic Versioning and Conventional Commits. This repo uses Earthly.
To contribute please:
- Fork the repository
- Create a feature branch
- Submit a pull request
Prior to submitting a pull request, please run:
earthly +test
One can run the benchmark test from the bench package:
make bench
to run the benchmarkmake bench-stats
to see the benchmark stats