Skip to content

Limits and gaps

Each item says what the code does. Where an item rests on the Artemis grammar rather than a run, it says so. The core's own limits are in Limitations.

Where the tested ground ends

  • One broker. Apache ActiveMQ Artemis, and only apache/artemis:2.57.0-alpine. No other AMQP 1.0 broker has been run, and Artemis is the only Management that ships. A Management for another broker is possible by design and untested.
  • One version. The code comment says Artemis 2.44.0 or later, but the suite ran 2.44.0 until the fixture moved to 2.57.0 and no longer does. Nothing between has been run.
  • The pin does not advance by itself. The version lives in a Go constant, which no automatic updater reads yet: see the open pin ticket.
  • One configuration. The suite runs against broker.xml with delivery-count persistence and journal sync on. The stock configuration is covered only by the capability it declares.

Conditions on what it promises

  • Survival is the broker's configuration. SurvivesRestart is true only when the broker persists the delivery count, persistence is on and the journal syncs non-transactional writes. On a stock Artemis it is false, and AtLeastOnce then needs AllowVolatile.
  • One attempt can be lost to an orderly restart. RestartLosesLastFailure: the failure after the last delivery is held in memory, so a restart returns the message numbered one lower. The count does not reset. A graceful stop does the same to a message in hand in about half of trials on 2.44.0 and 2.57.0, because Artemis counts the cancelled delivery back down and persists it. A SIGKILL did not lose the count in 44 trials. See the report on what rolls Artemis's delivery count back.
  • The operator owns the topology. Nothing durable is created, and the checks are made at subscribe, not continuously. A queue destroyed later ends the source: the adapter logs the receiver is gone.

Not offered

  • Replay. StartAt: AllAvailable is refused at construction.
  • Delayed redelivery. A Backoff or MaxNackDelay is refused at construction. A decline is handed back at once.
  • Request and reply. A Responder is refused; Request returns ErrNotSupported.
  • Overflow policies. Credit is the bound; nothing is shed. A requested shed policy degrades to that.
  • Non-durable publishing. Every publish is durable and waits for the broker's disposition.
  • TLS options. amqps:// is dialled with no TLS configuration from Settings, and no other connection option is exposed.

Behaviour to know about

  • Duplicates are dropped silently. The adapter keys Artemis's duplicate-id cache with a hash of Source and ID. A second publish of the same pair is accepted and not enqueued, and Publish returns nil. The pair is what CloudEvents defines uniqueness by, so this is deliberate, and a surprise when a test reuses an ID.
  • The unrouted check is once per address per session. A subscriber that appears afterwards is the next session's news, and one that goes away is not noticed.
  • A pattern's > becomes #, which also matches zero words. The core's > needs at least one token, and Artemis's # matches zero or more, so a subscription on a.> has the broker address a.#, which also matches the address a. This follows from the grammar the code encodes in wildcardMatches; it has not been run. If a subject exactly equal to the pattern's prefix is published, the subscription may receive it where the core's Match would say it should not.
  • The lease counts time in the adapter's buffer. The timer starts when a message is read off the link, so a message queued behind a slow handler can have its lease expire before it is delivered.
  • Only data sections are read. A message whose body is an amqp-value or amqp-sequence section has an empty body, and is probably unreadable.
  • Only string application properties are read. Properties of other types are ignored. When a format writes a header with several values, only the first is sent.
  • A queue's filter is not checked. Describe reads a queue's Filter, but verification does not compare it, so a durable multicast queue with a selector passes.
  • At-most-once competing queues linger. The group's queue, made by the broker, outlives its last member until the broker's auto-delete scan, about sixty seconds according to the test fixture.
  • Fixed timeouts. Attaching a link, settling and closing take at most 10 seconds each, and a management request at most 10 seconds. They are not settings.

Open

  • The pin ticket: the fixture's Artemis version is not tracked by an updater.