Models

Persistence and retention

Persistence = does the broker store the message? Retention = how long does the broker keep the stored message?

Model	Persistence	Retention
Event Log	Persistent	Retain until time/size policy removes it
Message Queue	Usually persistent	Retain until acknowledged or expired
Durable Pub/Sub	Persistent	Retain until subscriber state/retention policy allows removal
Ephemeral Pub/Sub	Not persistent	Do not retain for offline subscribers

ACK semantics

Model	What ACK means
Event Log	Consumer finished up to a position; message remains
Message Queue	Consumer finished the message; broker can remove it
Durable Pub/Sub	Subscriber finished the message; broker records progress
Ephemeral Pub/Sub	Usually no durable ACK state

Delivery guarantees

Guarantee	Meaning
At-most-once	Message may be lost, but is not redelivered
At-least-once	Message is not normally lost, but may be delivered more than once

(Solution: Consumers should usually be idempotent.) | | Exactly-once | Processing effect happens once under specific system guarantees |

Model	Common default
Event Log	At-least-once
Message Queue	At-least-once
Durable Pub/Sub	At-least-once
Ephemeral Pub/Sub	At-most-once / best effort

Ordering

Model	Typical ordering guarantee
Event Log	Usually partition-level ordering
Message Queue	Usually queue-level or limited ordering
Durable Pub/Sub	Usually limited or key-based ordering
Ephemeral Pub/Sub	Usually weak or best-effort ordering

Note that distributed messaging systems usually guarantee ordering within some scope, not globally.

Replay

Read a message again after it was already successfully processed. Majorly, this is the Even Log concept.

Model	Replay old messages?
Event Log	Yes, this is the core feature of event logs.
Message Queue	Usually no
Durable Pub/Sub	Sometimes, depending on retention and subscription state
Ephemeral Pub/Sub	No

Consumer group (same group)

One logical handler scales horizontally. Each message is assigned to one worker in that group. Majorly, this is the Even Log and Message Queue concept.

Model	Supported?
Event Log	Yes, common
Message Queue	Yes, common
Durable Pub/Sub	Usually yes
Ephemeral Pub/Sub	Depends on product

Most production systems offer a message to one consumer in a consumer group per delivery attempt, but cannot generally guarantee exactly-once business processing. A message may be redelivered after a timeout, crash, or failed acknowledgement. Consumers use idempotency, deduplication, transactional outbox/inbox patterns, and careful acknowledgement timing to make repeated deliveries safe.