Go channels for distributed queueing and event-driven systems
Today's wrong abstractions lead to complexity on maintainability in the future. Usage of synchronous interfaces to reflect asynchronous nature of messaging queues is a good example of inaccurate abstraction. Usage of pure Go channels is a proper solution to distills asynchronous semantic of queueing systems into the idiomatic native Golang code. The library adapts Go channels for various systems and interface. Please let us know via GitHub issues your needs about queuing technologies.
The library encourages developers to use Golang struct for asynchronous communication with peers. It helps engineers to define domain models, write correct, maintainable code. This library (swarm
) uses generic programming style to abstract queueing systems into the idiomatic Golang channels chan<- T
and <-chan T
. See the design pattern Golang channels for distributed event-driven architecture to learn philosophy and use-cases:
- readability: application uses pure Go code instead of vendor specific interfaces (learning time)
- portability: application is portable between various queuing systems or event brokers in same manner as sockets abstracts networking stacks (exchange queueing transport "on-the-fly" to resolve evolution of requirements)
- testability: unit testing focuses on pure biz logic, simplify dependency injections and mocking (pure unit tests).
- distribution: idiomatic architecture to build distributed topologies and scale-out Golang applications (clustering).
- serverless: of-the-shelf portable patterns for serverless applications (infrastructure as a code, aws cdk).
The library requires Go 1.18 or later due to usage of generics.
The latest version of the library is available at main
branch of this repository. All development, including new features and bug fixes, take place on the main
branch using forking and pull requests as described in contribution guidelines. The stable version is available via Golang modules.
Use go get
to retrieve the library and add it as dependency to your application.
go get -u github.com/fogfish/swarm
Example below is most simplest illustration of enqueuing and dequeuing message from AWS SQS.
package main
import (
"log/slog"
"github.com/fogfish/swarm"
"github.com/fogfish/swarm/broker/sqs"
"github.com/fogfish/swarm/enqueue"
"github.com/fogfish/swarm/dequeue"
)
func main() {
// create broker for AWS SQS
q, err := sqs.New("aws-sqs-queue-name")
if err != nil {
slog.Error("sqs broker has failed", "err", err)
return
}
// create Golang channels
rcv, ack := dequeue.Typed[string](q)
out := swarm.LogDeadLetters(enqueue.Typed[string](q))
// use Golang channels for I/O
go func() {
for msg := range rcv {
out <- msg.Object
ack <- msg
}
}()
q.Await()
}
Check the design pattern Distributed event-driven Golang channels for deep-dive into library philosophy. Also note, each supported broker comes with runnable examples that shows the library.
The following code snippet shows a typical flow of producing the messages using the library.
import (
"github.com/fogfish/swarm/broker/sqs"
"github.com/fogfish/swarm/enqueue"
)
// Use pure Golang struct to define semantic of messages and events
type Note struct {
ID string `json:"id"`
Text string `json:"text"`
}
// Spawn a new instance of the messaging broker
q, err := sqs.New("name-of-the-queue"), /* config options */)
// creates pair Golang channels dedicated for publishing
// messages of type [Note] through the messaging broker. The first channel
// is dedicated to emit messages. The second one is the dead letter queue that
// contains failed transmissions.
enq, dlq := enqueue.Typed[Note](q)
// Enqueue message of type Note
enq <- Note{ID: "note", Text: "some text"}
// Close the broker and release all resources
q.Close()
The following code snippet shows a typical flow of consuming the messages using the library.
import (
"github.com/fogfish/swarm/broker/sqs"
"github.com/fogfish/swarm/dequeue"
)
// Use pure Golang struct to define semantic of messages and events
type Note struct {
ID string `json:"id"`
Text string `json:"text"`
}
// Spawn a new instance of the messaging broker
q, err := sqs.New("name-of-the-queue", /* config options */)
// Create pair Golang channels dedicated for consuming
// messages of type Note from the messaging broker. The first channel
// is dedicated to receive messages. The second one is the channel to
// acknowledge consumption
deq, ack := dequeue.Typed[Note](q)
// consume messages and then acknowledge it
go func() {
for msg := range deq {
/* ... do something with msg.Object and ack the message ...*/
ack <- msg
}
}()
// Await messages from the broker
q.Await()
The library uses "option pattern" for the configuration, which is divided into two parts: a generic I/O kernel configuration and a broker-specific configuration. Please note that each configuration option is prefixed with With
and implemented in config.go files.
q, err := sqs.New("name-of-the-queue",
// WithXXX performs broker configuration
sqs.WithBatchSize(5),
// WithConfig performs generic kernel configuration
sqs.WithConfig(
swarm.WithSource("name-of-my-component"),
swarm.WithRetryConstant(10 * time.Millisecond, 3),
swarm.WithPollFrequency(10 * time.Second),
/* ... */
),
)
Usage of Golang channels as an abstraction raises a concern about grade of service on the message delivery guarantees. The library ensures exactly same grade of service as the underlying queueing system or event broker. Messages are delivered according to the promise once they are accepted by the remote side of queuing system. The library's built-in retry logic protects losses from temporary unavailability of the remote peer. However, Golang channels function as sophisticated "in-memory buffers," which can introduce a delay of a few milliseconds between scheduling a message to the channel and dispatching it to the remote peer. To handle catastrophic failures, choose one of the following policies to either accept or safeguard in-flight messages from potential loss.
At Most Once is best effort policy, where a message is published without any formal acknowledgement of receipt, and it isn't replayed. Some messages can be lost as subscribers are not required to acknowledge receipt.
The library implements asymmetric approaches for message handling. In the enqueue path, buffered Golang channels are used for both message emission and managing dead-letter queues. Similarly, the dequeue path uses buffered Golang channels to deliver messages to the consumer.
// Spawn a new instance of the messaging broker using At Most Once policy.
// The policy defines the capacity of Golang channel.
q, err := sqs.New("name-of-the-queue",
swarm.WithPolicyAtMostOnce(1000),
)
// for compatibility reasons two channels are returned on the enqueue path but
// dead-letter-queue is nil
enq, dlq := enqueue.Typed[Note](q)
// for compatibility reasons two channels are returned on the dequeue path but
// ack channel acts as /dev/null discards any sent message
deq, ack := dequeue.Typed[Note](q)
At Least Once is the default policy used by the library. The policy assume usage of "acknowledgement" protocol, which guarantees a message will be re-sent until it is formally acknowledged by a recipient. Messages should never be lost but it might be delivered more than once causing duplicate work to consumer.
The library also implements asymmetric approaches for message handling. In the enqueue path, unbuffered Golang channels are used to emit messages and manage the dead-letter queue, resulting in a delayed guarantee. This means that enqueuing additional messages is blocked until the dead-letter queue is fully resolved. Alternatively, the application can opt for a synchronous protocol to enqueue messages.
In the dequeue path, buffered Golang channels are used to deliver messages to the consumer and acknowledge their processing. While consumer acknowledgment ensures reliable message delivery, it may lead to message duplication.
// Spawn a new instance of the messaging broker using At Least Once policy.
// At Least Once policy is the default one, no needs to explicitly declare it.
// Use it only if you need to define other capacity for dequeue channel than
// the default one, which creates unbuffered channel
q, err := sqs.New("name-of-the-queue",
swarm.WithPolicyAtLeastOnce(1000),
)
// both channels are unbuffered
enq, dlq := enqueue.Typed[Note](q)
// buffered channels of capacity n
deq, ack := dequeue.Typed[Note](q)
Exactly Once is not supported by the library yet.
Usage of At Least Once policy (unbuffered channels) provides the delayed guarantee for producers. Let's consider the following example. If queue broker fails to send message A
then the channel enq
is blocked at sending message B
until the program consumes message A
from the dead-letter queue channel.
enq, dlq := enqueue.Typed[User](q)
enq <- User{ID: "A", Text: "some text by A"} // failed to send
enq <- User{ID: "B", Text: "some text by B"} // blocked until dlq is processed
enq <- User{ID: "C", Text: "some text by C"}
The delayed guarantee is efficient on batch processing, pipelining but might cause complication at transactional processing. Therefore, the library also support a synchronous variant to producing a message:
// Creates "synchronous" variant of the queue
user := enqueue.NewTyped[User](q)
// Synchronously enqueue the message. It ensure that message is scheduled for
// delivery to remote peer once function successfully returns.
if err := user.Enq(context.Background(), &User{ID: "A", Text: "some text by A"}); err != nil {
// handle error
}
The library guarantee ordering of the messages when they are produced over same Golang channel. Let's consider a following example:
user, _ := enqueue.Typed[User](q)
note, _ := enqueue.Typed[Note](q)
user <- &User{ID: "A", Text: "some text by A"}
note <- &Note{ID: "B", Text: "some note A"}
user <- &User{ID: "C", Text: "some text by A"}
The library guarantees following clauses A before C
and C after A
because both messages are produced to single channel user
. It do not guarantee clauses A before B
, B before C
or C after B
because multiple channels are used.
The library does not provide any higher guarantee than underlying message broker. For example, using SQS would not guarantee any ordering while SQS FIFO makes sure that messages of same type is ordered.
The library support slices of bytes []byte
as message type. It opens an opportunity for the many encoding options like JSON, Gob, etc.
import (
queue "github.com/fogfish/swarm/queue/bytes"
)
enq, dlq := enqueue.Bytes(q, "Note")
deq, ack := enqueue.Bytes(q, "Note")
Please see example about binary producer and consumer.
An event represents an immutable fact placed into the queuing system. It is conceptually similar to the Action defined by schema.org.
An action performed by a direct agent and indirect participants upon a direct object.
This type facilitates the development of event-driven solutions that treat data as a collection of immutable facts, which can be queried and processed in real-time. These applications process a logical event log, where each event represents a change to the current state of an object, such as which attributes were inserted, updated, or deleted (essentially a diff). Each event uniquely identifies the affected object using a unique identifier.
Unlike other solutions, this approach does not use an envelope for events. Instead, it pairs metadata and data side by side, making it more extendable.
type Meta struct {
swarm.Meta
About string `json:"about"`
}
type User struct {
ID string `json:"id"`
Text string `json:"text"`
}
// creates Golang channels to produce / consume messages
enq, dlq := enqueue.Event[Meta, User](q)
deq, ack := enqueue.Event[Meta, User](q)
Please see example about event producer and consumer.
The error handling on channel level is governed either by dead-letter queue or acknowledge protocol. The library provides swarm.WithStdErr
configuration option to pass the side channel to consume global errors. Use it as top level error handler.
stderr := make(chan error)
q, err := sqs.New("swarm-test",
sqs.WithConfig(
swarm.WithStdErr(stderr),
),
)
for err := range stderr {
// error handling loop
}
The existing message routing architecture assumes that a micro-batch of messages is read from the broker, dispatched to channels, and then waits for acknowledgments. A new micro-batch is not read until all messages are acknowledged, or the TimeToFlight
timer expires. In time-critical systems or serverless applications, a "fail fast" strategy is more effective (e.g., a Lambda function doesn't need to idle until the timeout).
Send negative acknowledgement to ack
channel to indicate error on message processing.
deq, ack := dequeue.Typed[Note](q)
// consume messages and then acknowledge it
for msg := range deq {
// negative ack on the error
if err := doSomething(msg.Object); err != nil {
ack <- msg.Fail(err)
continue
}
ack <- msg
}
The library primarily support development of serverless event-driven application using AWS service. The library provides AWS CDK Golang constructs to spawn consumers. See example of serverless consumer and corresponding AWS CDK application.
It consistently implements a pattern - "create Broker, attach Sinks".
package main
import (
"github.com/fogfish/scud"
"github.com/fogfish/swarm/broker/eventbridge"
)
func main() {
app := awscdk.NewApp(nil)
stack := awscdk.NewStack(app, jsii.String("swarm-example-eventbridge"),
&awscdk.StackProps{
Env: &awscdk.Environment{
Account: jsii.String(os.Getenv("CDK_DEFAULT_ACCOUNT")),
Region: jsii.String(os.Getenv("CDK_DEFAULT_REGION")),
},
},
)
// create broker
broker := eventbridge.NewBroker(stack, jsii.String("Broker"), nil)
broker.NewEventBus(nil)
broker.NewSink(
&eventbridge.SinkProps{
Source: []string{"swarm-example-eventbridge"},
Function: &scud.FunctionGoProps{
SourceCodeModule: "github.com/fogfish/swarm/broker/eventbridge",
SourceCodeLambda: "examples/dequeue/typed",
},
},
)
app.Synth(nil)
}
In a serverless environment, performing dequeue and enqueue operations can lead to race conditions. Specifically, the dequeue loop may complete before other emitted messages are processed.
rcv, ack := dequeue.Typed[/* .. */](broker1)
snd, dlq := enqueue.Typed[/* .. */](broker2)
for msg := range rcv {
snd <- // ...
// The ack would cause sleep of function in serverless.
// snd channel might not be flushed before function sleep.
// The library does not provide yet ultimate solution.
ack <- msg
}
Unfortunately, the library does not provide yet ultimate solution. Either sleep of sync senders are required.
The library is Apache Version 2.0 licensed and accepts contributions via GitHub pull requests:
- Fork it
- Create your feature branch (
git checkout -b my-new-feature
) - Commit your changes (
git commit -am 'Added some feature'
) - Push to the branch (
git push origin my-new-feature
) - Create new Pull Request
The build and testing process requires Go version 1.16 or later.
build and test library.
git clone https://github.com/fogfish/swarm
cd swarm
go test ./...
The commit message helps us to write a good release note, speed-up review process. The message should address two question what changed and why. The project follows the template defined by chapter Contributing to a Project of Git book.
If you experience any issues with the library, please let us know via GitHub issues. We appreciate detailed and accurate reports that help us to identity and replicate the issue.
TBD