Struct RabbitMqDelivery 

pub struct RabbitMqDelivery { /* private fields */ }

One RabbitMQ message, decoded into an Envelope.

Exactly one of ack, retry, defer or dead_letter must be called; the trait consumes the delivery so the compiler enforces “at most once”, and an un-acked delivery is redelivered by the broker when the consumer channel closes.

retry, defer and dead_letter publish before they ack, and propagate the publish error without acking, so a failure leaves the original message unacknowledged for the broker to redeliver rather than dropping the job. Because publishes are mandatory, a missing q.dead counts as a failure. A hold queue cannot be missing: retry and defer declare it themselves, immediately before publishing, but that declaration can be refused, and retry says what happens then.

When declare_dead_letter_queues is false this backend does not own q.dead, so dead_letter rejects the message (requeue = false) instead of publishing to it, and logs why.

Trait Implementations

impl Debug for RabbitMqDelivery

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Delivery for RabbitMqDelivery

fn retry<'async_trait>( self: Box<Self>, next: Envelope, delay: Duration, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>

where Self: 'async_trait,

Publish first, ack second.

next is written into the hold queue for delay (declared on the spot, mandatory, waited on for a publisher confirm) and only once the broker has taken responsibility for it is the original acked. If the publish fails the ? returns before the ack, so the original stays unacknowledged and the broker redelivers it, so the job is retried rather than silently dropped. The reverse order would lose a job on any broker hiccup between the two.

The delay is rounded up to RabbitMqOptions::retry_granularity, and next returns to q at the priority it carries, 0 after Envelope::next_attempt, so it joins the back of the queue.

Three things count as that failure, and all three leave the original unacked (the worker counts a settle failure and the broker redelivers the job when the consumer channel closes):

• the hold queue exists with different arguments, so the declaration is refused with PRECONDITION_FAILED. The declaration runs on its own channel, so concurrent publishes are untouched;
• next.queue was never declared through this backend, so the hold queue’s durability and dead-letter target are unknown (Error::UnknownQueue);
• delay is longer than MAX_DEFERRAL_MS (~24.8 days), which is refused rather than shortened, because a hold never releases a job early.

fn defer<'async_trait>( self: Box<Self>, next: Envelope, delay: Duration, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>

where Self: 'async_trait,

Publish first, ack second, the same rule and the same failure modes as retry.

The delay is rounded up to RabbitMqOptions::deferred_granularity instead, and next carries its queue’s top priority, so it returns ahead of the backlog.

A caveat on “ahead of the backlog”: a consumer with prefetch N is already holding up to N messages of that backlog, and the returning deferral cannot overtake those. It is first among what is still on the queue.

fn envelope(&self) -> &Envelope

The message this delivery carries.

fn ack<'async_trait>( self: Box<Self>, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>

where Self: 'async_trait,

Successfully processed; remove from broker.

fn dead_letter<'life0, 'async_trait>( self: Box<Self>, reason: &'life0 str, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Failed permanently (or attempts exhausted); route to dead-letter storage.