Messages can be delivered unsuccessfully (e.g. if the transacted session used to consume them is rolled back). Such a message goes back to its queue ready to be redelivered. However, this means it is possible for a message to be delivered again and again without success thus remaining in the queue indefinitely, clogging the system.
There are 2 ways to deal with these undelivered messages:
-
Delayed redelivery.
It is possible to delay messages redelivery. This gives the client some time to recover from any transient failures and to prevent overloading its network or CPU resources.
-
Dead Letter Address.
It is also possible to configure a dead letter address so that after a specified number of unsuccessful deliveries, messages are removed from their queue and sent to the dead letter address. These messages will not be delivered again from this queue.
Both options can be combined for maximum flexibility.
1. Delayed Redelivery
Delaying redelivery can often be useful in cases where clients regularly fail or rollback. Without a delayed redelivery, the system can get into a "thrashing" state, with delivery being attempted, the client rolling back, and delivery being re-attempted ad infinitum in quick succession, consuming valuable CPU and network resources.
#Persist Redelivery
Two Journal update records are stored every time a redelivery happens. One for the number of deliveries that happened, and one in case a scheduled redelivery is being used.
It is recommended to keep max-redelivery-records=1 in situations where you are operating with very short redelivery delays as you will be creating unecessary records on the journal.
1.1. Configuring Delayed Redelivery
Delayed redelivery is defined in the address-setting configuration:
<!-- delay redelivery of messages for 5s -->
<address-setting match="exampleQueue">
<!-- default is 1.0 -->
<redelivery-delay-multiplier>1.5</redelivery-delay-multiplier>
<!-- default is 0 (no delay) -->
<redelivery-delay>5000</redelivery-delay>
<!-- default is 0.0) -->
<redelivery-collision-avoidance-factor>0.15</redelivery-collision-avoidance-factor>
<!-- default is redelivery-delay * 10 -->
<max-redelivery-delay>50000</max-redelivery-delay>
</address-setting>
If a redelivery-delay
is specified, Apache ActiveMQ Artemis will wait this delay before redelivering the messages.
By default, there is no redelivery delay (redelivery-delay
is set to 0).
Other subsequent messages will be delivery regularly, only the cancelled message will be sent asynchronously back to the queue after the delay.
You can specify a multiplier (the redelivery-delay-multiplier
) that will take effect on top of the redelivery-delay
.
Each time a message is redelivered the delay period will be equal to the previous delay * redelivery-delay-multiplier
.
A max-redelivery-delay
can be set to prevent the delay from becoming too large.
The max-redelivery-delay
is defaulted to redelivery-delay
* 10.
Example:
-
redelivery-delay=5000, redelivery-delay-multiplier=2, max-redelivery-delay=15000, redelivery-collision-avoidance-factor=0.0
-
Delivery Attempt 1. (Unsuccessful)
-
Wait Delay Period: 5000
-
Delivery Attempt 2. (Unsuccessful)
-
Wait Delay Period: 10000 // (5000 * 2) < max-delay-period. Use 10000
-
Delivery Attempt 3: (Unsuccessful)
-
Wait Delay Period: 15000 // (10000 * 2) > max-delay-period: Use max-delay-delivery
-
Address wildcards can be used to configure redelivery delay for a set of addresses (see Understanding the Wildcard Syntax), so you don’t have to specify redelivery delay individually for each address.
The redelivery-delay
can be also be modified by configuring the redelivery-collision-avoidance-factor
.
This factor will be made either positive or negative at random to control whether the ultimate value will increase or decrease the redelivery-delay
.
Then it’s multiplied by a random number between 0.0 and 1.0.
This result is then multiplied by the redelivery-delay
and then added to the redelivery-delay
to arrive at the final value.
The algorithm may sound complicated but the bottom line is quite simple: the larger redelivery-collision-avoidance-factor
you choose the larger the variance of the redelivery-delay
will be.
The redelivery-collision-avoidance-factor
must be between 0.0 and 1.0.
Example:
-
redelivery-delay=1000, redelivery-delay-multiplier=1, max-redelivery-delay=15000, redelivery-collision-avoidance-factor=0.5, (bold values chosen using
java.util.Random
)-
Delivery Attempt 1. (Unsuccessful)
-
Wait Delay Period: 875 // 1000 + (1000 * ((0.5 * -1) * .25)
-
Delivery Attempt 2. (Unsuccessful)
-
Wait Delay Period: 1375 // 1000 + (1000 * ((0.5 * 1) * .75)
-
Delivery Attempt 3: (Unsuccessful)
-
Wait Delay Period: 975 // 1000 + (1000 * ((0.5 * -1) * .05)
-
This feature can be particularly useful in environments where there are multiple consumers on the same queue all interacting transactionally with the same external system (e.g. a database). If there is overlapping data in messages which are consumed concurrently then one transaction can succeed while all the rest fail. If those failed messages are redelivered at the same time then this process where one consumer succeeds and the rest fail will continue. By randomly padding the redelivery-delay by a small, configurable amount these redelivery "collisions" can be avoided.
1.2. Example
See the examples chapter for an example which shows how delayed redelivery is configured and used with JMS.
2. Dead Letter Addresses
To prevent a client infinitely receiving the same undelivered message (regardless of what is causing the unsuccessful deliveries), messaging systems define dead letter addresses: after a specified unsuccessful delivery attempts, the message is removed from its queue and sent to a dead letter address.
Any such messages can then be diverted to queue(s) where they can later be perused by the system administrator for action to be taken.
Apache ActiveMQ Artemis’s addresses can be assigned a dead letter address. Once the messages have been unsuccessfully delivered for a given number of attempts, they are removed from their queue and sent to the relevant dead letter address. These dead letter messages can later be consumed from the dead letter address for further inspection.
2.1. Configuring Dead Letter Addresses
Dead letter address is defined in the address-setting configuration:
<!-- undelivered messages in exampleQueue will be sent to the dead letter address
deadLetterQueue after 3 unsuccessful delivery attempts -->
<address-setting match="exampleQueue">
<dead-letter-address>deadLetterAddress</dead-letter-address>
<max-delivery-attempts>3</max-delivery-attempts>
</address-setting>
If a dead-letter-address
is not specified, messages will be removed after max-delivery-attempts
unsuccessful attempts.
By default, messages are redelivered 10 times at the maximum.
Set max-delivery-attempts
to -1 for infinite redeliveries.
A dead letter address
can be set globally for a set of matching addresses and you can set max-delivery-attempts
to -1 for a specific address setting to allow infinite redeliveries only for this address.
Address wildcards can be used to configure dead letter settings for a set of addresses (see Understanding the Wildcard Syntax).
2.2. Dead Letter Properties
Dead letter messages get special properties.
2.3. Automatically Creating Dead Letter Resources
It’s common to segregate undelivered messages by their original address.
For example, a message sent to the stocks
address that couldn’t be delivered for some reason might be ultimately routed to the DLQ.stocks
queue, and likewise a message sent to the orders
address that couldn’t be delivered might be routed to the DLQ.orders
queue.
Using this pattern can make it easy to track and administrate undelivered messages.
However, it can pose a challenge in environments which predominantly use auto-created addresses and queues.
Typically administrators in those environments don’t want to manually create an address-setting
to configure the dead-letter-address
much less the actual address
and queue
to hold the undelivered messages.
The solution to this problem is to set the auto-create-dead-letter-resources
address-setting
to true
(it’s false
by default) so that the broker will create the address
and queue
to deal with the undelivered messages automatically.
The address
created will be the one defined by the dead-letter-address
.
A MULTICAST
queue
will be created on that address
.
It will be named by the address
to which the message was previously sent, and it will have a filter defined using the property _AMQ_ORIG_ADDRESS
so that it will only receive messages sent to the relevant address
.
The queue
name can be configured with a prefix and suffix.
See the relevant settings in the table below:
address-setting |
default |
---|---|
|
|
|
(empty string) |
Here is an example configuration:
<address-setting match="#">
<dead-letter-address>DLA</dead-letter-address>
<max-delivery-attempts>3</max-delivery-attempts>
<auto-create-dead-letter-resources>true</auto-create-dead-letter-resources>
<dead-letter-queue-prefix></dead-letter-queue-prefix> <!-- override the default -->
<dead-letter-queue-suffix>.DLQ</dead-letter-queue-suffix>
</address-setting>
The queue holding the undeliverable messages can be accessed directly either by using the queue’s name by itself (e.g. when using the core client) or by using the fully qualified queue name (e.g. when using a JMS client) just like any other queue.
Also, note that the queue is auto-created which means it will be auto-deleted as per the relevant address-settings
.
2.4. Example
See: Dead Letter section of the Examples for an example that shows how dead letter resources can be statically configured and used with JMS.
3. Delivery Count Persistence
In normal use, Apache ActiveMQ Artemis does not update delivery count persistently until a message is rolled back (i.e. the delivery count is not updated before the message is delivered to the consumer). In most messaging use cases, the messages are consumed, acknowledged and forgotten as soon as they are consumed. In these cases, updating the delivery count persistently before delivering the message would add an extra persistent step for each message delivered, implying a significant performance penalty.
However, if the delivery count is not updated persistently before the message delivery happens, in the event of a server crash, messages might have been delivered but that will not have been reflected in the delivery count.
During the recovery phase, the server will not have knowledge of that and will deliver the message with redelivered
set to false
while it should be true
.
As this behavior breaks strict JMS semantics, Apache ActiveMQ Artemis allows to persist delivery count before message delivery but this feature is disabled by default due to performance implications.
To enable it, set persist-delivery-count-before-delivery
to true
in broker.xml
:
<persist-delivery-count-before-delivery>true</persist-delivery-count-before-delivery>