Skip to content

Capabilities

Source: Capabilities() in backend.go. The values below were read from a live Capabilities() call against Artemis 2.57.0, once with the test fixture's broker.xml and once with the image's stock configuration. What each field means is the core's: see Capabilities.

Field Value Decided by
DeliveryModes AtMostOnce, AtLeastOnce fixed
Patterns true fixed
SurvivesRestart Features.PersistsDeliveryCount && Features.Persists the broker's configuration, read at New
RestartLosesLastFailure Features.PersistsDeliveryCount the broker's configuration, read at New
ExtendsLease true fixed
Ordered true fixed
RoutesAroundDegraded true fixed
Overflow none (nil) fixed
Replays false fixed
RestartCanSpendUnseenAttempt false not set
DelaysRedelivery false (MaxRedeliveryDelay zero) not set
RequestReply false not set

The two broker-dependent values are constant after New. They are read once, not re-read.

The two that depend on the broker

On Artemis, amqp.Artemis{}.Features reads three attributes of the broker management resource:

Attribute Feeds
PersistDeliveryCountBeforeDelivery PersistsDeliveryCount
PersistenceEnabled and JournalSyncNonTransactional, both Persists

An attribute the broker does not expose reads as false: a feature that cannot be verified is not declared.

Broker configuration SurvivesRestart RestartLosesLastFailure
the test fixture's broker.xml (persist-delivery-count-before-delivery, persistence-enabled and journal-sync-non-transactional all true) true true
the image's stock configuration, apache/artemis:2.57.0-alpine false false

On a broker declaring SurvivesRestart: false, an AtLeastOnce subscription is refused at construction unless it sets AllowVolatile. See handle a degraded construction.

Why it declares RestartLosesLastFailure

Artemis writes the delivery count as it delivers, not as an attempt fails. The failure after the last delivery is held in memory until the next delivery, so an orderly restart returns the message numbered one lower than the attempts it has had. The budget is still bounded and the count does not reset.

A graceful stop does the same to a message in hand: Artemis's shutdown cancels the consumer, counts the cancelled delivery back down and persists that. The ticket records it in about half of trials, on both 2.44.0 and 2.57.0, with no broker setting that prevents it. A kill with SIGKILL, by contrast, lost the count in none of 44 trials. See the report on what rolls Artemis's delivery count back and the closed ticket messaging-amqp#1.

The gaps, plainly

  • No replay. Replays: false. StartAt: AllAvailable is refused at construction. A queue delivers what it holds.
  • No delayed redelivery. DelaysRedelivery: false. A subscription with a Backoff or a MaxNackDelay is refused at construction on this backend. A decline is handed back at once.
  • No request and reply. RequestReply: false. A Responder subscription is refused at construction, and Request returns ErrNotSupported, which is a degraded bus, not a refused one.
  • No overflow policies. Overflow is empty, so a requested ShedNewest or ShedOldest degrades to the backend's native behaviour, which is credit: nothing is shed here. See what happens at a full queue and Delivery, settlement and flow control.
  • No declared restart-attempt loss. RestartCanSpendUnseenAttempt: false: the adapter does not declare that a restart can spend an attempt no process saw.
  • Survival is conditional. See above: it is the broker's configuration, not the adapter's.