xtr_amqp_messenger ¶
A RabbitMQ transport for xtr-messenger, built on its taskiq bridge.
Installed beside xtr-messenger — or as xtr-messenger[amqp] — an amqp://
or amqps:// DSN is served by :class:AmqpTransportFactory, found by its
scheme with no wiring: a connection and its queues, a retry ladder, delays
through the delayed-message exchange, and real dead-lettering.
AmqpTransportFactory ¶
Bases: TransportFactoryInterface, WorkerProvidingInterface, PublisherClosingInterface
Builds amqp:// transports, sharing one broker per DSN.
Apply these to everything it builds.
serializer is what both halves of every transport it builds write
and read messages with. Producer and consumer are separate processes,
so build both from the same configuration: the guarantee that matters
is that what one side writes, the other reads back exactly.
create ¶
Build one broker for the group, and a sender per named queue.
close_publishers
async
¶
Close every connection this factory's senders opened on first publish.
A worker's broker is closed by its own run loop and was never recorded, so this releases the producer side only.
worker ¶
worker(group: Mapping[str, TransportConfig], bus: MessageBusInterface, *, event_dispatcher: EventDispatcherInterface | None = None, resetter: AsyncResetter | None = None) -> WorkerInterface
Build a runnable worker consuming only group.
Every declared message is registered on the broker as a task that
dispatches into bus, so the result needs no further wiring —
await worker.run() and it consumes, with bus deciding which
handlers run. It listens on exactly these queues, so a worker serving
a different transport is unaffected.
Import the modules that declare your messages and handlers before
calling this; a message that has not been declared cannot be bound.
It handles at most max_async_tasks messages at once.
event_dispatcher hears the worker events, each message reported
as received from the transport whose queue it arrived on.
resetter has reset() awaited after each message, settled
either way, so long-lived services are cleared between units of work.
With a dead-letter queue configured, a delivery that can never be handled — unparsable, naming a task nobody registered, or whose body does not decode into its message — is quarantined there at once rather than retried.
Raises:
-
MixedDsnError–If the transports do not share one connection.
DeadLetterMiddleware ¶
Bases: TaskiqMiddleware
Republishes a message to the dead-letter queue once retries run out.
Dead-letter to queue_name after max_attempts deliveries.
on_error
async
¶
Dead-letter the message if this was its last attempt.
Reliability
dataclass
¶
Reliability(max_attempts: int = 3, base_delay_seconds: float = 1.0, dead_letter_queue: str | None = DEFAULT_DEAD_LETTER_QUEUE)
How hard to try, and where a message goes when trying stops.
Attributes:
-
max_attempts(int) –Total deliveries including the first, not retries on top of it.
-
base_delay_seconds(float) –Delay before the first retry; later retries back off from it.
-
dead_letter_queue(str | None) –Where a message is republished once attempts run out. Set to
Noneto drop exhausted messages instead.
from_settings
classmethod
¶
Read reliability out of a transport's settings.
Anything the settings do not mention keeps its value from
defaults, so a factory constructed with a policy in code can
still have one setting overridden per transport.
Raises:
-
InvalidTransportOptionError–If a value is present but not a number.
create_amqp_broker ¶
create_amqp_broker(url: str, options: AmqpOptions | None = None, queues: Sequence[str] = (), clock: ClockInterface | None = None) -> AioPikaBroker
Build a RabbitMQ broker.
Construction opens no connection, so this is safe to call at import time — which is what a module-level broker in a worker entrypoint needs.
Pass queues to publish to more than one queue; a
:class:~xtr_messenger.bridge.taskiq.taskiq_sender.TaskiqSender then
selects one by name, which is how a routing table maps message types onto
queues.
A delayed message is held back as options.delay says — natively on
RabbitMQ 4.4 and later, in a delay queue of its own on an older one, by
default; clock reads the time a native delay ends at. With
auto_setup off nothing is declared but delay queues, the dead-letter
queue and the plugin's delayed exchange included: the topology must
already exist, or startup fails naming what is missing.
declared_queues ¶
Return the queues broker declares, which are the ones it consumes.
Useful at worker startup to log the workload a process has taken on, and to confirm a deployment scoped its workers the way it intended.