kafka_go is an abstraction over popular kafka client sarama (https://github.com/Shopify/sarama). Though sarama provides good enough APIs to integrate with a kafka cluster but still lags simplicity and needs a bit of domain knowledge even for a standard use case.
End-user has to maintain fail-safety, reclaim after re-balancing or similar scenarios, API doesn't seem very intuitive for the first time kafka users. kafka_go tries to solve all such problems with its easy to understand APIs to start consuming from a kafka cluster with significant less domain knowledge and complete fail-safety.
The package abstracts out fairly easy and intuitive abstraction to integrate with a kafka cluster with minimal effort. It takes care of the complexity of fail-safety, possible leaks scenarios, re-balance handling, and export easy to use functional interface. The API is divided into Consumer and Producer modules -
The consumer is an important abstraction of any PubSub library. The package tries to make it as simple as possible.
type Consumer interface {
//Start should trigger the actual message consumption process, it should be blocking in nature to avoid killing
//process immaturely.
Start(ctx context.Context)
//Stop should trigger the closure of the consumption process. Should cancel the context to relieve resources and take
//care of possible leaks
Stop()
}
The package implements this simple generic Consumer interface for Kafka. A user can instantiate a consumer instance with minimal effort, Example -
consumer, err := NewKafkaConsumer(&KafkaConsumerParam{
Brokers: []string{"broker-1", "broker-2", "broker-3"},
GroupID: "test-cg",
OffsetInitial: OtNewest,
Topics: []string{"test-topic"},
Handlers: map[string]TopicHandler{
"test-topic": &testTopicHandler{},
},
})
if err != nil {
Logger.Panicf("Error creating consumer instance %v", err)
}
consumer.Start(context.Background())
Refer to examples in the code base for better understanding.
It provides a single place to configure the Consumer client. ConsumerParam provider rich set of keys configurable by the user with fairly standard defaults to start consuming out of the box.
type KafkaConsumerParam struct {
// Brokers in kafka clusters
Brokers []string
// Consumer group id of this consumer group
GroupID string
// List of topics to start listening from
Topics []string
// Topic to handlers map to consumer message from a topic.
Handlers map[string]TopicHandler
// [Optional]
// Topic to its fallback handler map.
// If the Main handler returns false, it will try to fallback handler.
// it will commit the offset not matter fallback handler returns true or false.
// default - "no fallback"
Fallbacks map[string]TopicHandler
// [Optional]
// List of middleware to be triggered post claim of every message & before actual
// message handling. Middleware will be triggered in increasing order order of index.
// default - "no middleware"
middleware []ConsumerMiddleware
// [Optional]
// List of interceptor, like Middleware it trigger post claim of every message, but unlike
// middleware interceptor is available after the actual handler return. Interceptors are
// triggered in layered manner, lower index being the outer layer and vice versa. This is
// similar to recursive call, the one called first will return last.
// default - "noOpInterceptor"
interceptor []ConsumerInterceptor
// [Optional]
// Attach a meta map with every claimed message before passing it to actual handler, can be used
// to persist state during the lifecycle of a claimed message. Middleware or Interceptor can also
// use this meta to store variable across. default - "false"
MessageMeta bool
// [Optional]
// Client identity for logging purpose
ClientID string
// [Optional]
// The initial offset to use if no offset was previously committed.
// Should be OtNewest or OtOldest. defaults - OtNewest.
OffsetInitial OffsetType
// [Optional]
// kafka cluster version. eg - "2.2.1" default - "2.3.0"
// supports versions from "0.8.x" to "2.3.x"
Version string
}
type TopicHandler interface {
// Handle gets the actual message to be handled. Business logic for a given
// message should ideally be implemented here.
Handle(ctx context.Context, message *SubscriberMessage) bool
}
TopicHandler is an intuitive interface that exposes a single function Handle. An implementation of this interface should be used as handler for messages from a topic. Technically the business logic post receiving the message from a topic should trigger from here. The handler has not been tightly bound to a topic on purpose to allow it to be used for multiple topics at once. The implementation should be thread/goroutine safe obviously. Example -
// test topic handler
type testTopicHandler struct {
}
func (t *testTopicHandler) Handle(ctx context.Context, msg *SubscriberMessage) bool {
Logger.Printf("Topic: %v, Partition: %v, SubscriberMessage: %v", msg.Topic, msg.Partition, string(msg.Value))
return true
}
type ConsumerMiddleware func(ctx context.Context, msg *SubscriberMessage)
Middleware is a construct similar to Before advice in AOP. Middleware comes just after claiming a message from a topic and before passing it to its handler down the line. Middleware can be used to decorate the message or preprocess something across topics before actual handling. This saves from writing common code across all the topic handlers. The middleware can be optionally used to enrich message meta which can be used inside actual handler or later in the handling pipeline. Multiple independent middleware can be used which pre-process the message in a pre-defined manner. Use cases can be -
- Filtering messages to pass relevant messages to the actual handler
- Validating message for correctness
- logging, throttling, decorating etc are few other possibilities
type ConsumerInterceptor func(ctx context.Context, msg *SubscriberMessage, handler func(context.Context, *SubscriberMessage) bool) bool
Interceptor is a construct similar to Around advice in AOP. Interceptor life cycle start just after claiming a message around the handler and available even after the handler return. Interceptor might come handy start some processing before actual message handling and ending it after the handling is done. Use cases can be -
- Newrelic transaction start before message handling and close after handler returns
- Collect data around message processing like time to process a message or so
- Take some common action on expected/unexpected handling of a message, like publish to dead letter
Multiple independent interceptor can be used which works in a layered fashion. For example if the List of interceptor consist of IC_1, IC_2, IC_3, the processing will look like -
IC_1
{
before_1
IC_2
{
before_2
IC_3
{
before_3
message_handler(.....)
after_3
}
after_2
}
after_1
}
The next obvious abstraction that any PubSub library will export is Producer.
type Producer interface {
// PublishSync sends a message to the PubSub cluster in Sync way. Call to this function is blocking and
// returns only after the publish is done or result in an error. Meta contains the publish related meta info
PublishSync(message *PublisherMessage) (meta map[string]interface{}, err error)
// PublishSyncBulk sends messages to the PubSub cluster in sync all at once. Call to this function is blocking
// and return only after publish attempt is done for all the messages. Return error if the bulk publish is
// partially successful
PublishSyncBulk(messages []*PublisherMessage) error
// PublishAsync sends a message to the PubSub cluster in Async way. Call to this function is non-blocking and
// returns immediately.
PublishAsync(message *PublisherMessage)
// PublishAsyncBulk sends messages in bulk to the PubSub cluster in Async way. Call to this function is non-blocking
// and return immediately
PublishAsyncBulk(messages []*PublisherMessage)
// Close triggers the closure of the associated producer client to avoid any leaks
Close()
}
The package implements the Producer interface to provide a rich set of functionality to produce messages to kafka cluster in every possible way. The abstraction clearly hides the complexity to manage channels, provides high throughput async bulk publish functionality to publish a fairly huge amount of messages at once without blocking the main process. Though the user is advised to take care of CPU/Memory implications of producing too many messages at once. Example -
producer, err := NewKafkaProducerQuick([]string{"broker-1", "broker-2", "broker-3"})
if err != nil {
Logger.Fatalf("Error creating producer client, %v", err)
}
data, _ := json.Marshal("test message")
producer.PublishAsync(&PublisherMessage{
Topic: "test-topic",
Key: "test-key",
Data: data,
})
For better understanding and implementation, refer to the coe base and test examples.
Package kept the Consumer and Producer interfaces fairly generic and can be implemented for different PubSub frameworks like Redis-PubSub, AWS SNS-SQS, Rabbit-MQ, etc in future. The intent of this package is to make integration to Kafka like PubSub framework fairly easy with minimal efforts. The author decided to create this package after personally struggling making the integration stable and fail-safe. The author observed that most of the developers are struggling around the same problem and end up writing their own abstraction to make it look clean. Although most of these abstractions are similar but mostly written considering their own repo structure under consideration. This package tried taking all such abstraction under consideration and export fairly reusable code across various use cases.
The license has been chosen on purpose. Feel free to take fork and raise a pull request
for any possible bugs or new use cases. Always keep in mind the intent of the package to keep it fairly simple to understand and use.