Queue Subscriptions Compared to Bookmark Replays

At first glance, a subscription to a queue can look very similar to a bookmark replay. In both cases, messages are provided in order from the transaction log. In both cases, a message that is not processed can be retrieved and redelivered. In both cases, AMPS allows a subscription to exactly specify the messages of interest using content filtering.

There are a few main differences:

• Delivery Model

  With a bookmark subscription, a message from a given topic can be delivered to any number of subscribers and processed multiple times. A given subscriber can replay the same messages any number of times, if needed.

  With a queue subscription, the server guarantees that once a message is processed, it is not delivered from that topic again.

• Delivery Limits

  With a bookmark subscription, by default AMPS will deliver messages to the subscription as fast as the physical hardware, disk, and network allow.

  With a queue subscription, AMPS intentionally limits the number of messages delivered to a subscriber, to help ensure that messages are processed in a timely manner. Since each message should be delivered only once, delivering every message to a single subscriber would run the risk of having a slow consumer holding messages that it is unable to process, while a faster consumer sits idle.

• Acknowledgment

  With a bookmark subscription, the application and client libraries are responsible for tracking which messages have been processed and providing the correct recovery point. There is no state stored in the AMPS server as to which messages have been consumed by a given subscription.

  With a queue subscription, the AMPS server tracks whether a given message has been processed. A queue subscriber must notify the server that it is finished with a given message and is ready for more work. The AMPS client libraries contain support for making this process easy, as well as the ability to batch acknowledgments to provide high throughput while maintaining efficient processing and the guarantees for queues.