JetStream Model Deep Dive

Stream Limits, Retention, and Policy

Streams store data on disk, but we cannot store all data forever so we need ways to control their size automatically.

There are 3 features that come into play when Streams decide how long they store data.

The Retention Policy describes based on what criteria a set will evict messages from its storage:

Retention PolicyDescription


Limits are set for how many messages, how big the storage and how old messages may be.


Messages are kept until they are consumed: meaning delivered ( by the consumer filtering on the message's subject (in this mode of operation you can not have any overlapping consumers defined on the Stream - each subject captured by the stream can only have one consumer at a time)) to a subscribing application and explicitly acknowledged by that application.


Messages are kept as long as there are Consumers on the stream (matching the message's subject if they are filtered consumers) for which the message has not yet been ACKed. Once all currently defined consumers have received explicit acknowledgement from a subscribing application for the message it is then removed from the stream.

In all Retention Policies the basic limits apply as upper bounds, these are MaxMsgs for how many messages are kept in total, MaxBytes for how big the set can be in total and MaxAge for what is the oldest message that will be kept. These are the only limits in play with LimitsPolicy retention.

One can then define additional ways a message may be removed from the Stream earlier than these limits. In WorkQueuePolicy the messages will be removed as soon as the Consumer received an Acknowledgement. In InterestPolicy messages will be removed as soon as all Consumers of the stream for that subject have received an Acknowledgement for the message.

In both WorkQueuePolicy and InterestPolicy the age, size and count limits will still apply as upper bounds.

A final control is the Maximum Size any single message may have. NATS have it's own limit for maximum size (1 MiB by default), but you can say a Stream will only accept messages up to 1024 bytes using MaxMsgSize.

The Discard Policy sets how messages are discarded when limits set by LimitsPolicy are reached. The DiscardOld option removes old messages making space for new, while DiscardNew refuses any new messages.

The WorkQueuePolicy mode is a specialized mode where a message, once consumed and acknowledged, is removed from the Stream.

Message Deduplication

JetStream support idempotent message writes by ignoring duplicate messages as indicated by the Nats-Msg-Id header.

nats req -H Nats-Msg-Id:1 hello1
nats req -H Nats-Msg-Id:1 hello2
nats req -H Nats-Msg-Id:1 hello3
nats req -H Nats-Msg-Id:1 hello4

Here we set a Nats-Msg-Id:1 header which tells JetStream to ensure we do not have duplicates of this message - we only consult the message ID not the body.

nats stream info ORDERS

and in the output you can see that the duplicate publications were detected and only one message (the first one) is actually stored in the stream


            Messages: 1
               Bytes: 67 B

The default window to track duplicates in is 2 minutes, this can be set on the command line using --dupe-window when creating a stream, though we would caution against large windows.

Acknowledgement Models

Streams support acknowledging receiving a message, if you send a Request() to a subject covered by the configuration of the Stream the service will reply to you once it stored the message. If you just publish, it will not. A Stream can be set to disable Acknowledgements by setting NoAck to true in it's configuration.

Consumers have 3 acknowledgement modes:



This requires every message to be specifically acknowledged, it's the only supported option for pull-based Consumers


In this mode if you acknowledge message 100 it will also acknowledge message 1-99, this is good for processing batches and to reduce ack overhead


No acknowledgements are supported

To understand how Consumers track messages we will start with a clean ORDERS Stream and DISPATCH Consumer.

nats str info ORDERS

            Messages: 0
               Bytes: 0 B
            FirstSeq: 0
             LastSeq: 0
    Active Consumers: 1

The Set is entirely empty

nats con info ORDERS DISPATCH

  Last Delivered Message: Consumer sequence: 1 Stream sequence: 1
    Acknowledgment floor: Consumer sequence: 0 Stream sequence: 0
        Pending Messages: 0
    Redelivered Messages: 0

The Consumer has no messages outstanding and has never had any (Consumer sequence is 1).

We publish one message to the Stream and see that the Stream received it:

nats pub ORDERS.processed "order 4"
Published 7 bytes to ORDERS.processed
$ nats str info ORDERS

            Messages: 1
               Bytes: 53 B
            FirstSeq: 1
             LastSeq: 1
    Active Consumers: 1

As the Consumer is pull-based, we can fetch the message, ack it, and check the Consumer state:

nats con next ORDERS DISPATCH
--- received on ORDERS.processed
order 4

Acknowledged message

$ nats con info ORDERS DISPATCH

  Last Delivered Message: Consumer sequence: 2 Stream sequence: 2
    Acknowledgment floor: Consumer sequence: 1 Stream sequence: 1
        Pending Messages: 0
    Redelivered Messages: 0

The message got delivered and acknowledged - Acknowledgement floor is 1 and 1, the sequence of the Consumer is 2 which means its had only the one message through and got acked. Since it was acked, nothing is pending or redelivering.

We'll publish another message, fetch it but not Ack it this time and see the status:

nats pub ORDERS.processed "order 5"
Published 7 bytes to ORDERS.processed

Get the next message from the consumer (but do not acknowledge it)

nats consumer next ORDERS DISPATCH --no-ack
--- received on ORDERS.processed
order 5

Show the consumer info

nats consumer info ORDERS DISPATCH

  Last Delivered Message: Consumer sequence: 3 Stream sequence: 3
    Acknowledgment floor: Consumer sequence: 1 Stream sequence: 1
        Pending Messages: 1
    Redelivered Messages: 0

Now we can see the Consumer has processed 2 messages (obs sequence is 3, next message will be 3) but the Ack floor is still 1 - thus 1 message is pending acknowledgement. Indeed this is confirmed in the Pending messages.

If I fetch it again and again do not ack it:

nats consumer next ORDERS DISPATCH --no-ack
--- received on ORDERS.processed
order 5

Show the consumer info again

nats consumer info ORDERS DISPATCH

  Last Delivered Message: Consumer sequence: 4 Stream sequence: 3
    Acknowledgment floor: Consumer sequence: 1 Stream sequence: 1
        Pending Messages: 1
    Redelivered Messages: 1

The Consumer sequence increases - each delivery attempt increases the sequence - and our redelivered count also goes up.

Finally, if I then fetch it again and ack it this time:

nats consumer next ORDERS DISPATCH 
--- received on ORDERS.processed
order 5

Acknowledged message

Show the consumer info

nats consumer info ORDERS DISPATCH

  Last Delivered Message: Consumer sequence: 5 Stream sequence: 3
    Acknowledgment floor: Consumer sequence: 1 Stream sequence: 1
        Pending Messages: 0
    Redelivered Messages: 0

Having now Acked the message there are no more pending.

Additionally, there are a few types of acknowledgements:



nil, +ACK

Acknowledges a message was completely handled



Signals that the message will not be processed now and processing can move onto the next message, NAK'd message will be retried



When sent before the AckWait period indicates that work is ongoing and the period should be extended by another equal to AckWait



Acknowledges the message was handled and requests delivery of the next message to the reply subject. Only applies to Pull-mode.



Instructs the server to stop redelivery of a message without acknowledging it as successfully processed

So far all of the examples were the AckAck type of acknowledgement, by replying to the Ack with the body as indicated in Bytes you can pick what mode of acknowledgement you want.

All of these acknowledgement modes, except AckNext, support double acknowledgement - if you set a reply subject when acknowledging the server will in turn acknowledge having received your ACK.

The +NXT acknowledgement can have a few formats: +NXT 10 requests 10 messages and +NXT {"no_wait": true} which is the same data that can be sent in a Pull Request.

Exactly Once Semantics

JetStream supports Exactly Once publication and consumption by combining Message Deduplication and double acks.

On the publishing side you can avoid duplicate message ingestion using the Message Deduplication feature.

Consumers can be 100% sure a message was correctly processed by requesting the server Acknowledge having received your acknowledgement (sometimes referred to as double-acking) by calling the message's AckSync() (rather than Ack()) function which sets a reply subject on the Ack and waits for a response from the server on the reception and processing of the acknowledgement. If the response received from the server indicates success you can be sure that the message will never be re-delivered by the consumer (due to a loss of your acknowledgement).

Consumer Starting Position

When setting up a Consumer you can decide where to start, the system supports the following for the DeliverPolicy:



Delivers all messages that are available


Delivers the latest message, like a tail -n 1 -f


Delivers only new messages that arrive after subscribe time


Delivers from a specific time onward. Requires OptStartTime to be set


Delivers from a specific stream sequence. Requires OptStartSeq to be set

Regardless of what mode you set, this is only the starting point. Once started it will always give you what you have not seen or acknowledged. So this is merely how it picks the very first message.

Let's look at each of these, first we make a new Stream ORDERS and add 100 messages to it.

Now create a DeliverAll pull-based Consumer:

nats consumer add ORDERS ALL --pull --filter ORDERS.processed --ack none --replay instant --deliver all 
nats consumer next ORDERS ALL
--- received on ORDERS.processed
order 1

Acknowledged message

Now create a DeliverLast pull-based Consumer:

nats consumer add ORDERS LAST --pull --filter ORDERS.processed --ack none --replay instant --deliver last
nats consumer next ORDERS LAST
--- received on ORDERS.processed
order 100

Acknowledged message

Now create a MsgSetSeq pull-based Consumer:

nats consumer add ORDERS TEN --pull --filter ORDERS.processed --ack none --replay instant --deliver 10
nats consumer next ORDERS TEN
--- received on ORDERS.processed
order 10

Acknowledged message

And finally a time-based Consumer. Let's add some messages a minute apart:

nats stream purge ORDERS
for i in 1 2 3
  nats pub ORDERS.processed "order ${i}"
  sleep 60

Then create a Consumer that starts 2 minutes ago:

nats consumer add ORDERS 2MIN --pull --filter ORDERS.processed --ack none --replay instant --deliver 2m
nats consumer next ORDERS 2MIN
--- received on ORDERS.processed
order 2

Acknowledged message

Ephemeral Consumers

So far, all the Consumers you have seen were Durable, meaning they exist even after you disconnect from JetStream. In our Orders scenario, though the MONITOR a Consumer could very well be a short-lived thing there just while an operator is debugging the system, there is no need to remember the last seen position if all you are doing is wanting to observe the real-time state.

In this case, we can make an Ephemeral Consumer by first subscribing to the delivery subject, then creating a durable and giving it no durable name. An Ephemeral Consumer exists as long as any subscription is active on its delivery subject. It is automatically be removed, after a short grace period to handle restarts, when there are no subscribers.

Terminal 1:

nats sub my.monitor

Terminal 2:

nats consumer add ORDERS --filter '' --ack none --target 'my.monitor' --deliver last --replay instant --ephemeral

The --ephemeral switch tells the system to make an Ephemeral Consumer.

Consumer Message Rates

Typically what you want is if a new Consumer is made the selected messages are delivered to you as quickly as possible. You might want to replay messages at the rate they arrived though, meaning if messages first arrived 1 minute apart and you make a new Consumer it will get the messages a minute apart.

This is useful in load testing scenarios etc. This is called the ReplayPolicy and have values of ReplayInstant and ReplayOriginal.

You can only set ReplayPolicy on push-based Consumers.

nats consumer add ORDERS REPLAY --target out.original --filter ORDERS.processed --ack none --deliver all --sample 100 --replay original
     Replay Policy: original

Now let's publish messages into the Set 10 seconds apart:

for i in 1 2 3                                                                                                                                                      <15:15:35
  nats pub ORDERS.processed "order ${i}"
  sleep 10
Published [ORDERS.processed] : 'order 1'
Published [ORDERS.processed] : 'order 2'
Published [ORDERS.processed] : 'order 3'

And when we consume them they will come to us 10 seconds apart:

nats sub -t out.original
Listening on [out.original]
2020/01/03 15:17:26 [#1] Received on [ORDERS.processed]: 'order 1'
2020/01/03 15:17:36 [#2] Received on [ORDERS.processed]: 'order 2'
2020/01/03 15:17:46 [#3] Received on [ORDERS.processed]: 'order 3'

Ack Sampling

In the earlier sections we saw that samples are being sent to a monitoring system. Let's look at that in depth; how the monitoring system works and what it contains.

As messages pass through a Consumer you'd be interested in knowing how many are being redelivered and how many times but also how long it takes for messages to be acknowledged.

Consumers can sample Ack'ed messages for you and publish samples so your monitoring system can observe the health of a Consumer. We will add support for this to NATS Surveyor.


You can configure a Consumer for sampling bypassing the --sample 80 option to nats consumer add, this tells the system to sample 80% of Acknowledgements.

When viewing info of a Consumer you can tell if it's sampled or not:

nats consumer info ORDERS NEW

Output contains

     Sampling Rate: 100

Storage Overhead

JetStream file storage is very efficient, storing as little extra information about the message as possible.

We do store some message data with each message, namely:

  • Message headers

  • The subject it was received on

  • The time it was received

  • The message payload

  • A hash of the message

  • The message sequence

  • A few other bits like the length of the subject and the length of headers

Without any headers the size is:

length of the message record (4bytes) + seq(8) + ts(8) + subj_len(2) + subj + msg + hash(8)

A 5 byte hello message without headers will take 39 bytes.

With headers:

length of the message record (4bytes) + seq(8) + ts(8) + subj_len(2) + subj + hdr_len(4) + hdr + msg + hash(8)

So if you are publishing many small messages the overhead will be, relatively speaking, quite large, but for larger messages the overhead is very small. If you publish many small messages it's worth trying to optimize the subject length.

Last updated