kafka-go

Motivations

We rely on both Go and Kafka a lot at Segment. Unfortunately, the state of the Go client libraries for Kafka at the time of this writing was not ideal. The available options were:

• sarama, which is by far the most popular but is quite difficult to work with. It is poorly documented, the API exposes low level concepts of the Kafka protocol, and it doesn't support recent Go features like contexts. It also passes all values as pointers which causes large numbers of dynamic memory allocations, more frequent garbage collections, and higher memory usage.

• confluent-kafka-go is a cgo based wrapper around librdkafka, which means it introduces a dependency to a C library on all Go code that uses the package. It has much better documentation than sarama but still lacks support for Go contexts.

• goka is a more recent Kafka client for Go which focuses on a specific usage pattern. It provides abstractions for using Kafka as a message passing bus between services rather than an ordered log of events, but this is not the typical use case of Kafka for us at Segment. The package also depends on sarama for all interactions with Kafka.

This is where kafka-go comes into play. It provides both low and high level APIs for interacting with Kafka, mirroring concepts and implementing interfaces of the Go standard library to make it easy to use and integrate with existing software.

Note:

In order to better align with our newly adopted Code of Conduct, the kafka-go project has renamed our default branch to main. For the full details of our Code Of Conduct see this document.

Kafka versions

kafka-go is currently tested with Kafka versions 0.10.1.0 to 2.7.1. While it should also be compatible with later versions, newer features available in the Kafka API may not yet be implemented in the client.

Go versions

kafka-go requires Go version 1.15 or later.

Connection

The Conn type is the core of the kafka-go package. It wraps around a raw network connection to expose a low-level API to a Kafka server.

Here are some examples showing typical use of a connection object:

// to produce messages
topic := "my-topic"
partition := 0

conn, err := kafka.DialLeader(context.Background(), "tcp", "localhost:9092", topic, partition)
if err != nil {
    log.Fatal("failed to dial leader:", err)
}

conn.SetWriteDeadline(time.Now().Add(10*time.Second))
_, err = conn.WriteMessages(
    kafka.Message{Value: []byte("one!")},
    kafka.Message{Value: []byte("two!")},
    kafka.Message{Value: []byte("three!")},
)
if err != nil {
    log.Fatal("failed to write messages:", err)
}

if err := conn.Close(); err != nil {
    log.Fatal("failed to close writer:", err)
}

// to consume messages
topic := "my-topic"
partition := 0

conn, err := kafka.DialLeader(context.Background(), "tcp", "localhost:9092", topic, partition)
if err != nil {
    log.Fatal("failed to dial leader:", err)
}

conn.SetReadDeadline(time.Now().Add(10*time.Second))
batch := conn.ReadBatch(10e3, 1e6) // fetch 10KB min, 1MB max

b := make([]byte, 10e3) // 10KB max per message
for {
    n, err := batch.Read(b)
    if err != nil {
        break
    }
    fmt.Println(string(b[:n]))
}

if err := batch.Close(); err != nil {
    log.Fatal("failed to close batch:", err)
}

if err := conn.Close(); err != nil {
    log.Fatal("failed to close connection:", err)
}

To Create Topics

By default kafka has the auto.create.topics.enable='true' (KAFKA_CFG_AUTO_CREATE_TOPICS_ENABLE='true' in the bitnami/kafka kafka docker image). If this value is set to 'true' then topics will be created as a side effect of kafka.DialLeader like so:

// to create topics when auto.create.topics.enable='true'
conn, err := kafka.DialLeader(context.Background(), "tcp", "localhost:9092", "my-topic", 0)
if err != nil {
    panic(err.Error())
}

If auto.create.topics.enable='false' then you will need to create topics explicitly like so:

// to create topics when auto.create.topics.enable='false'
topic := "my-topic"

conn, err := kafka.Dial("tcp", "localhost:9092")
if err != nil {
    panic(err.Error())
}
defer conn.Close()

controller, err := conn.Controller()
if err != nil {
    panic(err.Error())
}
var controllerConn *kafka.Conn
controllerConn, err = kafka.Dial("tcp", net.JoinHostPort(controller.Host, strconv.Itoa(controller.Port)))
if err != nil {
    panic(err.Error())
}
defer controllerConn.Close()


topicConfigs := []kafka.TopicConfig{
    {
        Topic:             topic,
        NumPartitions:     1,
        ReplicationFactor: 1,
    },
}

err = controllerConn.CreateTopics(topicConfigs...)
if err != nil {
    panic(err.Error())
}

To Connect To Leader Via a Non-leader Connection

// to connect to the kafka leader via an existing non-leader connection rather than using DialLeader
conn, err := kafka.Dial("tcp", "localhost:9092")
if err != nil {
    panic(err.Error())
}
defer conn.Close()
controller, err := conn.Controller()
if err != nil {
    panic(err.Error())
}
var connLeader *kafka.Conn
connLeader, err = kafka.Dial("tcp", net.JoinHostPort(controller.Host, strconv.Itoa(controller.Port)))
if err != nil {
    panic(err.Error())
}
defer connLeader.Close()

To list topics

conn, err := kafka.Dial("tcp", "localhost:9092")
if err != nil {
    panic(err.Error())
}
defer conn.Close()

partitions, err := conn.ReadPartitions()
if err != nil {
    panic(err.Error())
}

m := map[string]struct{}{}

for _, p := range partitions {
    m[p.Topic] = struct{}{}
}
for k := range m {
    fmt.Println(k)
}

Because it is low level, the Conn type turns out to be a great building block for higher level abstractions, like the Reader for example.

Reader

A Reader is another concept exposed by the kafka-go package, which intends to make it simpler to implement the typical use case of consuming from a single topic-partition pair. A Reader also automatically handles reconnections and offset management, and exposes an API that supports asynchronous cancellations and timeouts using Go contexts.

Note that it is important to call Close() on a Reader when a process exits. The kafka server needs a graceful disconnect to stop it from continuing to attempt to send messages to the connected clients. The given example will not call Close() if the process is terminated with SIGINT (ctrl-c at the shell) or SIGTERM (as docker stop or a kubernetes restart does). This can result in a delay when a new reader on the same topic connects (e.g. new process started or new container running). Use a signal.Notify handler to close the reader on process shutdown.

// make a new reader that consumes from topic-A, partition 0, at offset 42
r := kafka.NewReader(kafka.ReaderConfig{
    Brokers:   []string{"localhost:9092","localhost:9093", "localhost:9094"},
    Topic:     "topic-A",
    Partition: 0,
    MaxBytes:  10e6, // 10MB
})
r.SetOffset(42)

for {
    m, err := r.ReadMessage(context.Background())
    if err != nil {
        break
    }
    fmt.Printf("message at offset %d: %s = %s\n", m.Offset, string(m.Key), string(m.Value))
}

if err := r.Close(); err != nil {
    log.Fatal("failed to close reader:", err)
}

Consumer Groups

kafka-go also supports Kafka consumer groups including broker managed offsets. To enable consumer groups, simply specify the GroupID in the ReaderConfig.

ReadMessage automatically commits offsets when using consumer groups.

// make a new reader that consumes from topic-A
r := kafka.NewReader(kafka.ReaderConfig{
    Brokers:   []string{"localhost:9092", "localhost:9093", "localhost:9094"},
    GroupID:   "consumer-group-id",
    Topic:     "topic-A",
    MaxBytes:  10e6, // 10MB
})

for {
    m, err := r.ReadMessage(context.Background())
    if err != nil {
        break
    }
    fmt.Printf("message at topic/partition/offset %v/%v/%v: %s = %s\n", m.Topic, m.Partition, m.Offset, string(m.Key), string(m.Value))
}

if err := r.Close(); err != nil {
    log.Fatal("failed to close reader:", err)
}

There are a number of limitations when using consumer groups:

• (*Reader).SetOffset will return an error when GroupID is set
• (*Reader).Offset will always return -1 when GroupID is set
• (*Reader).Lag will always return -1 when GroupID is set
• (*Reader).ReadLag will return an error when GroupID is set
• (*Reader).Stats will return a partition of -1 when GroupID is set

Explicit Commits

kafka-go also supports explicit commits. Instead of calling ReadMessage, call FetchMessage followed by CommitMessages.

ctx := context.Background()
for {
    m, err := r.FetchMessage(ctx)
    if err != nil {
        break
    }
    fmt.Printf("message at topic/partition/offset %v/%v/%v: %s = %s\n", m.Topic, m.Partition, m.Offset, string(m.Key), string(m.Value))
    if err := r.CommitMessages(ctx, m); err != nil {
        log.Fatal("failed to commit messages:", err)
    }
}

When committing messages in consumer groups, the message with the highest offset for a given topic/partition determines the value of the committed offset for that partition. For example, if messages at offset 1, 2, and 3 of a single partition were retrieved by call to FetchMessage, calling CommitMessages with message offset 3 will also result in committing the messages at offsets 1 and 2 for that partition.

Managing Commits

By default, CommitMessages will synchronously commit offsets to Kafka. For improved performance, you can instead periodically commit offsets to Kafka by setting CommitInterval on the ReaderConfig.

// make a new reader that consumes from topic-A
r := kafka.NewReader(kafka.ReaderConfig{
    Brokers:        []string{"localhost:9092", "localhost:9093", "localhost:9094"},
    GroupID:        "consumer-group-id",
    Topic:          "topic-A",
    MaxBytes:       10e6, // 10MB
    CommitInterval: time.Second, // flushes commits to Kafka every second
})

Writer

To produce messages to Kafka, a program may use the low-level Conn API, but the package also provides a higher level Writer type which is more appropriate to use in most cases as it provides additional features:

• Automatic retries and reconnections on errors.
• Configurable distribution of messages across available partitions.
• Synchronous or asynchronous writes of messages to Kafka.
• Asynchronous cancellation using contexts.
• Flushing of pending messages on close to support graceful shutdowns.
• Creation of a missing topic before publishing a message. Note! it was the default behaviour up to the version v0.4.30.

// make a writer that produces to topic-A, using the least-bytes distribution
w := &kafka.Writer{
    Addr:     kafka.TCP("localhost:9092", "localhost:9093", "localhost:9094"),
    Topic:   "topic-A",
    Balancer: &kafka.LeastBytes{},
}

err := w.WriteMessages(context.Background(),
    kafka.Message{
        Key:   []byte("Key-A"),
        Value: []byte("Hello World!"),
    },
    kafka.Message{
        Key:   []byte("Key-B"),
        Value: []byte("One!"),
    },
    kafka.Message{
        Key:   []byte("Key-C"),
        Value: []byte("Two!"),
    },
)
if err != nil {
    log.Fatal("failed to write messages:", err)
}

if err := w.Close(); err != nil {
    log.Fatal("failed to close writer:", err)
}

Missing topic creation before publication

// Make a writer that publishes messages to topic-A.
// The topic will be created if it is missing.
w := &Writer{
    Addr:                   kafka.TCP("localhost:9092", "localhost:9093", "localhost:9094"),
    Topic:                  "topic-A",
    AllowAutoTopicCreation: true,
}

messages := []kafka.Message{
    {
        Key:   []byte("Key-A"),
        Value: []byte("Hello World!"),
    },
    {
        Key:   []byte("Key-B"),
        Value: []byte("One!"),
    },
    {
        Key:   []byte("Key-C"),
        Value: []byte("Two!"),
    },
}

var err error
const retries = 3
for i := 0; i < retries; i++ {
    ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
    defer cancel()

    // attempt to create topic prior to publishing the message
    err = w.WriteMessages(ctx, messages...)
    if errors.Is(err, kafka.LeaderNotAvailable) || errors.Is(err, context.DeadlineExceeded) {
        time.Sleep(time.Millisecond * 250)
        continue
    }

    if err != nil {
        log.Fatalf("unexpected error %v", err)
    }
    break
}

if err := w.Close(); err != nil {
    log.Fatal("failed to close writer:", err)
}

Writing to multiple topics

Normally, the WriterConfig.Topic is used to initialize a single-topic writer. By excluding that particular configuration, you are given the ability to define the topic on a per-message basis by setting Message.Topic.

```go w := &kafka.Writer{ Addr: kafka.TCP("localhost:9092", "localhost:9093", "localhost:9094"), // NOTE: When Topic is not defined here, each Message must define it instead. Bal