Consumer Details

Consumers are how client applications get the messages stored in the streams. You can have many consumers on a single stream. Consumers are like a view on a stream, can filter messages and have some state (maintained by the servers) associated with them.

Consumers can be 'durable' or 'ephemeral'.

Durable versus ephemeral consumers

Durable consumer persist message delivery progress on the server side. A durable consumer can be retrieved by name and shared between client instance for load balancing. It can be made highly available through replicas.

An ephemeral consumer does not persist delivery progress and will automatically be deleted when there are no more client instances connected.

Durable consumers

Durable consumers are meant to be used by multiple instances of an application, either to distribute and scale out the processing, or to persist the position of the consumer over the stream between runs of an application.

Durable consumers as the name implies are meant to last 'forever' and are typically created and deleted administratively rather than by the application code which only needs to specify the durable's well known name to use it.

You create a durable consumer using the nats consumer add CLI tool command, or programmatically by passing a durable name option to the subscription creation call.

Ephemeral consumers

Ephemeral consumers are meant to be used by a single instance of an application (e.g. to get its own replay of the messages in the stream).

Ephemeral consumers are not meant to last 'forever', they are defined automatically at subscription time by the client library and disappear after the application disconnect.

You (automatically) create an ephemeral consumer when you call the js.Subscribe function without specifying the Durable or Bind subscription options. Calling Drain on that subscription automatically deletes the underlying ephemeral consumer. You can also explicitly create an ephemeral consumer by not passing a durable name option to the jsm.AddConsumer call.

Ephemeral consumers otherwise have the same control over message acknowledged and re-delivery as durable consumers.

Push and Pull consumers

Clients implement two implementations of consumers identified as 'push' or 'pull'.

Push consumers

Push consumers receive messages on a specific subject where message flow is controlled by the server. Load balancing is supported through NATS core queue groups. The messages from the stream are distributed automatically between the subscribing clients to the push consumers.

Pull consumers

Pull consumers request messages explicitly from the server in batches, giving the client full control over dispatching, flow control, pending (unacknowledged) messages and load balancing. Pull consuming client make fetch() calls in a dispatch loop.

We recommend using pull consumers for new projects. In particular when scalability, detailed flow control or error handling are a design focus. Most client API have been updated to provide convenient interfaces for consuming messages through callback handler or iterators without the need to manage message retrieval.

fetch() calls can be immediate or have a defined timeout, allowing for either controlled (1 by 1) consumption or realtime delivery with minimal polling overhead.

Pull consumers create less CPU load on the NATS servers and therefore scale better (note that the push consumers are still quite fast and scalable, you may only notice the difference between the two if you have sustained high message rates).

Pull

func ExampleJetStream() {
    nc, err := nats.Connect("localhost")
    if err != nil {
        log.Fatal(err)
    }

	// Use the JetStream context to produce and consumer messages
	// that have been persisted.
	js, err := nc.JetStream(nats.PublishAsyncMaxPending(256))
	if err != nil {
		log.Fatal(err)
	}

	js.AddStream(&nats.StreamConfig{
		Name:     "FOO",
		Subjects: []string{"foo"},
	})

	js.Publish("foo", []byte("Hello JS!"))

	// Publish messages asynchronously.
	for i := 0; i < 500; i++ {
		js.PublishAsync("foo", []byte("Hello JS Async!"))
	}
	select {
	case <-js.PublishAsyncComplete():
	case <-time.After(5 * time.Second):
		fmt.Println("Did not resolve in time")
	}

	// Create Pull based consumer with maximum 128 inflight.
	sub, _ := js.PullSubscribe("foo", "wq", nats.PullMaxWaiting(128))

	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
	defer cancel()

	for {
		select {
		case <-ctx.Done():
			return
		default:
		}

        // Fetch will return as soon as any message is available rather than wait until the full batch size is available, using a batch size of more than 1 allows for higher throughput when needed.
		msgs, _ := sub.Fetch(10, nats.Context(ctx))
		for _, msg := range msgs {
			msg.Ack()
		}
	}
}

A push consumer can also be used in some other use cases such as without a queue group, or with no acknowledgement or cumulative acknowledgements.

Push

func ExampleJetStream() {
	nc, err := nats.Connect("localhost")
	if err != nil {
		log.Fatal(err)
	}

	// Use the JetStream context to produce and consumer messages
	// that have been persisted.
	js, err := nc.JetStream(nats.PublishAsyncMaxPending(256))
	if err != nil {
		log.Fatal(err)
	}

	js.AddStream(&nats.StreamConfig{
		Name:     "FOO",
		Subjects: []string{"foo"},
	})

	js.Publish("foo", []byte("Hello JS!"))

	// Publish messages asynchronously.
	for i := 0; i < 500; i++ {
		js.PublishAsync("foo", []byte("Hello JS Async!"))
	}
	select {
	case <-js.PublishAsyncComplete():
	case <-time.After(5 * time.Second):
		fmt.Println("Did not resolve in time")
	}

	// Create async consumer on subject 'foo'. Async subscribers
	// ack a message once exiting the callback.
	js.Subscribe("foo", func(msg *nats.Msg) {
		meta, _ := msg.Metadata()
		fmt.Printf("Stream Sequence  : %v\n", meta.Sequence.Stream)
		fmt.Printf("Consumer Sequence: %v\n", meta.Sequence.Consumer)
	})

	// Async subscriber with manual acks.
	js.Subscribe("foo", func(msg *nats.Msg) {
		msg.Ack()
	}, nats.ManualAck())

	// Async queue subscription where members load balance the
	// received messages together.
	// If no consumer name is specified, either with nats.Bind()
	// or nats.Durable() options, the queue name is used as the
	// durable name (that is, as if you were passing the
	// nats.Durable(<queue group name>) option.
	// It is recommended to use nats.Bind() or nats.Durable()
	// and preferably create the JetStream consumer beforehand
	// (using js.AddConsumer) so that the JS consumer is not
	// deleted on an Unsubscribe() or Drain() when the member
	// that created the consumer goes away first.
	// Check Godoc for the QueueSubscribe() API for more details.
	js.QueueSubscribe("foo", "group", func(msg *nats.Msg) {
		msg.Ack()
	}, nats.ManualAck())

	// Subscriber to consume messages synchronously.
	sub, _ := js.SubscribeSync("foo")
	msg, _ := sub.NextMsg(2 * time.Second)
	msg.Ack()

	// We can add a member to the group, with this member using
	// the synchronous version of the QueueSubscribe.
	sub, _ = js.QueueSubscribeSync("foo", "group")
	msg, _ = sub.NextMsg(2 * time.Second)
	msg.Ack()

	// ChanSubscribe
	msgCh := make(chan *nats.Msg, 8192)
	sub, _ = js.ChanSubscribe("foo", msgCh)

	select {
	case msg := <-msgCh:
		fmt.Println("[Received]", msg)
	case <-time.After(1 * time.Second):
	}
}

Ordered Consumers

Ordered consumers are a convenient form of ephemeral push consumer for applications, that want to efficiently consume a stream for data inspection or analysis.

The API consumer is guaranteed delivery of messages in sequence and without gaps.

  • Always ephemeral - minimal overhead for the server

  • Single threaded in sequence dispatching

  • Client checks message sequence and will prevent gaps in the delivery

  • Can recover from server node failure and reconnect

  • Does not recover from client failure as it is ephemeral

func ExampleJetStream() {
	nc, err := nats.Connect("localhost")
	if err != nil {
		log.Fatal(err)
	}

	// Use the JetStream context to produce and consumer messages
	// that have been persisted.
	js, err := nc.JetStream(nats.PublishAsyncMaxPending(256))
	if err != nil {
		log.Fatal(err)
	}

	js.AddStream(&nats.StreamConfig{
		Name:     "FOO",
		Subjects: []string{"foo"},
	})

	js.Publish("foo", []byte("Hello JS!"))

	// ordered push consumer
	js.Subscribe("foo", func(msg *nats.Msg) {
		meta, _ := msg.Metadata()
		fmt.Printf("Stream Sequence  : %v\n", meta.Sequence.Stream)
		fmt.Printf("Consumer Sequence: %v\n", meta.Sequence.Consumer)
	}, nats.OrderedConsumer())
}

Delivery reliability

JetStream consumers can ensure not just the reliability of message delivery but also the reliability of the processing of the messages, even in the face of client application or downstream failures. It does so by using message level acknowledgements and message re-deliveries.

Consumers have an Acknowledgement Policy specifying the level of reliability required. In increasing order of reliability the available policies are: 'none' for no application level acknowledgements, 'all' where acknowledging a specific message also implicitly acknowledges all previous messages in the stream, and 'explicit' where each message must be individually acknowledged.

When the consumer is set to require explicit acknowledgements the client applications are able to use more than one kind of acknowledgement to indicate successful (or not) reception and processing of the messages being received from the consumer.

Applications can:

  • Acknowledge the successfull processing of a message (Ack()).

  • Acknowledge the successfull processing of a message and request an acknowledgement of the reception of the acknowledgement by the consumer (AckSync()).

  • Indicate that the processing is still in progress and more time is needed (inProgress()).

  • Negatively acknowledge a message, indicating that the client application is currently (temporarily) unable to process the message and that the consumer should attempt to re-deliver it (Nak()).

  • Terminate a message (typically, because there is a problem with the data inside the message such that the client application is never going to be able to process it), indicating that the consumer should not attempt to re-deliver the message (Term()).

After a message is sent from the consumer to a subscribing client application by the server an 'AckWait' timer is started. This timer is deleted when either a positive (Ack()) or a termination (Term()) acknowledgement is received from the client application. The timer gets reset upon reception of an in-progress (inProgress()) acknowledgement.

If at the end of a period of time no acknowledgement has been received from the client application, the server will attempt to re-deliver the message. If there is more than one client application instance subscribing to the consumer, there is no guarantee that the re-delivery would be to any particular client instance.

You can control the timing of re-deliveries using either the single AckWait duration attribute of the consumer, or as a sequence of durations in the BackOff attribute (which overrides AckWait).

You can also control the timing of re-deliveries when messages are negatively acknowledged with Nak(), by passing a nakDelay() option (or using NakWithDelay()), otherwise the re-delivery attempt will happen right after the reception of the Nak by the server.

"Dead Letter Queues" type functionality

You can set a maximum number of delivery attempts using the consumer's MaxDeliver setting.

Whenever a message reaches its maximum number of delivery attempts an advisory message is published on the $JS.EVENT.ADVISORY.CONSUMER.MAX_DELIVERIES.<STREAM>.<CONSUMER> subject. The advisory message's payload (use nats schema info io.nats.jetstream.advisory.v1.max_deliver for specific information) contains a stream_seq field that contains the sequence number of the message in the stream.

Similarly, whenever a client application terminates delivery attempts for the message using AckTerm an advisory message is published on the $JS.EVENT.ADVISORY.CONSUMER.MSG_TERMINATED.<STREAM>.<CONSUMER> subject, and its payload (see nats schema info io.nats.jetstream.advisory.v1.terminated) contains a stream_seq field.

You can leverage those advisory messages to implement "Dead Letter Queue" (DLQ) types of functionalities. For example:

  • If you only need to know about each time a message is 'dead' (considered un-re-deliverable by the consumer), then listening to the advisories is enough.

  • If you also need to have access to the message in question then you can use the message's sequence number included in the advisory to retrieve that specific message by sequence number from the stream. If a message reaches its maximum level of delivery attempts, it will still stay in the stream until it is manually deleted or manually acknowledged.

Last updated